XPENG has released video of its IRON humanoid following changing hand gestures without spoken commands. In the sequence a tester offers a handshake, switches to a fist bump, and moves into a high-five. When the target shifts, the robot brings in its other hand. The company attributes the behavior to an upgraded VLA 2.0 model running on-device, with stronger temporal understanding and future-behavior prediction. XPENG states that the robot reassesses the visual scene approximately every 0.2 seconds and predicts action changes within about six seconds. The footage is labeled an R&D test-version demonstration. No success rates, end-to-end latency figures, or results across multiple users have been published. The video shows continuous observation and adjustment rather than a fixed routine. IRON is XPENG’s general-purpose humanoid that uses the company’s Physical AI foundation. The same model family is used in the company’s vehicles.

Fixed gesture recognition followed by a complete pre-planned motion fails when a person changes intention midway. Continuous re-evaluation while the arm is already moving requires the model to update its plan in real time. XPENG’s description of on-device VLA 2.0 inference matches the visual result in the clip. The 0.2-second interval and six-second horizon are company figures. Independent timing measurements have not been released. The demonstration is useful as a qualitative illustration of adaptive upper-body interaction. It does not show the same behavior in unstructured public environments or with people who move more erratically than the staged tester. The company notes that final product behavior may differ from the development footage. No contact-force data or safety margins for the handshake and high-five interactions have been published. The sequence isolates the gesture-following loop under controlled conditions.

IRON incorporates a spine-like core, bionic muscle layout, and soft skin covering. Earlier XPENG videos have shown the robot walking, tracking speakers, and holding multi-language conversations. The current clip focuses on upper-body coordination and the ability to switch limbs when the target moves. The on-device claim indicates that inference occurs on the robot’s compute rather than in the cloud. IRON carries multiple Turing chips. The video does not show the internal compute configuration or power draw during the interaction. Continuous adjustment during a short social gesture is a limited but visible test of the loop. Broader claims of general social competence would require longer trials with more varied participants. No such data have been released. The demonstration is consistent with XPENG’s positioning of IRON as a platform that combines physical motion with real-time AI. The video remains the primary public evidence of the adaptive interaction behavior.

VLA 2.0 is XPENG’s vision-language-action model. The company has deployed versions in its vehicles and is applying an upgraded version to IRON. The video shows the robot switching hands and modifying its motion mid-interaction when the high-five position changes or a palm closes into a fist. The footage is a controlled development demonstration. No independent latency or success-rate data have been published. The demonstration focuses on adaptive upper-body interaction. Broader operational results in unstructured settings remain unreleased. The platform continues to be developed as a general-purpose humanoid using XPENG’s Physical AI stack. The sequence demonstrates on-device temporal reasoning applied to social motion. Final product behavior may differ from the R&D test version shown. No long-duration or multi-user trials of the same capability have been released.

XPENG’s IRON humanoid has been shown adjusting in real time during handshakes, fist bumps, and high-fives without spoken commands. The company attributes the behavior to an upgraded on-device VLA 2.0 model with temporal understanding and prediction. The robot is said to reassess the scene roughly every 0.2 seconds. The video demonstrates continuous observation and motion adjustment under controlled conditions. No independent metrics have been published. The sequence is an R&D test-version demonstration. The clip illustrates adaptive social interaction on the IRON platform. No long-duration or unstructured public trials of the same capability have been released. The demonstration focuses on upper-body coordination. Broader claims of general social competence remain unreleased. The platform continues development as a general-purpose humanoid.

UBTECH has started production at a factory in Liuzhou, Guangxi, that the company describes as designed for 10,000 industrial humanoid robots a year. The 14,000-square-meter plant, built with Siemens Digital Industries Software, uses flexible work cells rather than fixed conveyors. Cruzr robots handle depalletizing, parts transport, and replenishment alongside AGVs and unmanned forklifts. The company states that the line is timed for one completed robot roughly every ten minutes. Primary models are the Walker S and Cruzr series. Finished units undergo multi-hour stress testing and visual inspection before storage in a compact high-bay warehouse. The video shows robots and automated logistics inside the facility. No independent production-volume figures or sustained takt-time measurements have been published. The ten-minute interval is a company design target. Actual output over a full operating year has not been independently verified. The plant is presented as the first of its scale for industrial humanoids.

Separately, Unitree has fully open-sourced UnifoLM-WLA-1.0, a 6-billion-parameter embodied foundation model. The company states that a single model coordinates 64 tasks spanning tabletop manipulation and full-body mobile work, and generalizes across two-finger grippers and multiple five-finger dexterous hands. Training used approximately 2,500 hours of real-robot data. The model includes an embodied reasoner component and an action expert. Code, weights, and related datasets are available on GitHub and Hugging Face under an open license. The release followed a staged open-source plan that began in mid-September. No independent benchmarks of the 64-task set have been published outside company evaluations. The model is intended for developers working with Unitree platforms, particularly the G1. The open-source release lowers the barrier for fine-tuning on local data. Performance on customer robots remains dependent on further adaptation. The video for this brief focuses on the UBTECH factory; the Unitree model is a concurrent software announcement.

UBTECH’s Liuzhou facility is intended to move the company from lower-volume assembly toward higher-throughput manufacturing. The use of its own Cruzr robots for material handling is presented as “robots building robots.” Siemens software provides the manufacturing operations management layer, including work-order scheduling and serial-number traceability. The plant covers both bipedal Walker S units and wheeled Cruzr platforms. No customer order backlog or shipment figures tied specifically to the new line have been released. The video shows hanging Walker units, mobile robots, and the overall layout. It does not include a continuous multi-hour production run or a full accounting of human versus robotic labor on the line. For manufacturers evaluating humanoid production capacity, the relevant questions are the actual annual output, the yield rate after testing, and the cost structure at the stated volume. Those details are company-reported. The facility is an industrial-scale claim rather than a fully audited production result.

Unitree’s UnifoLM-WLA-1.0 is described as unifying tabletop and whole-body manipulation in one weight set. The embodied reasoner is built on a vision-language backbone and is trained on large volumes of spatial and multimodal samples. The action expert is trained on the real-robot hours. The company has released model weights, training code for action experts, and fine-tuning tools. Datasets referenced include Unitree open collections and other real-robot sources. The open-source step is notable because Unitree is a high-volume humanoid maker. Developers can now fine-tune the model on their own data rather than relying solely on published claims. No side-by-side comparison against other open foundation models on the same 64 tasks has been released. The model is a software release concurrent with the UBTECH factory news. Both announcements address scaling—one on the manufacturing side, one on the model side. Independent validation of either claim remains limited to company statements and early open-source materials.

UBTECH has begun production at a Liuzhou factory designed for 10,000 industrial humanoids a year, with a stated takt of one robot roughly every ten minutes. Cruzr robots and automated logistics handle much of the material movement. Separately, Unitree has open-sourced UnifoLM-WLA-1.0, a 6B model that the company says handles 64 tasks across tabletop and full-body work with different end-effectors. The factory video shows the plant layout and robotic material handling. No independent volume or takt measurements have been published. The model release includes code and weights. Performance on the full task set remains company-reported. Both developments address different aspects of scaling humanoid deployment. Actual sustained output and independent model benchmarks have not been released. The announcements are concurrent rather than linked. The factory is an industrial production claim. The model is an open-source software release.

Pollen Robotics’ Microduck, a 25 cm biped priced at $399 before taxes and shipping, has drawn substantial pre-order volume since its August 27 launch. The robot stands under 800 g, has 15 motors, a front camera, an 8×8 depth LiDAR, two IMUs, microphones, a speaker, NFC, Wi-Fi, and Bluetooth. Out of the box it can walk, sit, crouch, roller-skate, pick objects with its beak, and recover from many common falls. The open-source software stack allows developers to train new behaviors in simulation and transfer them to the robot. First deliveries were targeted before Christmas 2026 in North America, Europe, and the UK. After the initial rush the company stated it could no longer guarantee Christmas delivery for new orders and now estimates four to six months. The video shows the robot walking, falling, recovering, and performing trained behaviors. Demand figures reported by the company and third-party coverage exceed 10,000 units in the first week and later surpassed 20,000 after a public demonstration.

The combination of a low price, a small form factor, and an open reinforcement-learning pipeline has generated interest among makers and researchers. Microduck is not positioned as a finished household appliance. It is sold as a platform on which users can run supplied policies or train new ones. The software, including the SDK and training tools, is open source. Hardware designs are not fully open in the same way. The robot can be controlled with a game controller, physical accessories, and NFC-tagged objects. Multiple units can be used together for simple multi-robot experiments. The video demonstrates the default behaviors and the recovery sequence. No independent long-term reliability data from early customers have been published because the first units have not yet shipped in volume. The pre-order response indicates strong interest in an affordable, trainable biped. Delivery timelines have lengthened as order volume grew.

Pollen Robotics, part of Hugging Face, previously released Reachy Mini. Microduck adds legs and learned locomotion. The company has stated that default policies run locally on the robot. Developers can extend the stack and share new behaviors. The four colorways are Cream, Graphite, Lavender, and Sky. A game controller is included. The video shows the robot’s size relative to a desk and its ability to recover from falls. Search interest in the name rose sharply after the launch and again after a high-profile demonstration. The company has not published a final production capacity figure. Early reports placed first-week orders above 10,000 units. Later coverage indicated the total had moved past 20,000. These numbers are company and media reports rather than audited figures. The product remains in the pre-order phase. First shipments have not yet reached customers at scale.

Small, affordable robots that support on-device learned policies have been limited. Most research platforms are larger and more expensive. Microduck’s price point and open software stack lower the barrier for experimenting with reinforcement learning on physical hardware. The robot’s articulated beak can pick small objects. Its sensors support basic interaction and recovery. The video illustrates the walking gait, fall recovery, and object interaction. No formal safety certification for unsupervised use around children has been published. The company positions the robot for developers and enthusiasts rather than as a fully autonomous home companion. Demand has exceeded the initial delivery window. The company has adjusted its shipping estimate accordingly. The video remains the primary public demonstration of the robot’s out-of-box behaviors.

Pollen Robotics’ Microduck has attracted strong pre-order interest at a $399 price point. The 25 cm biped includes 15 motors, multiple sensors, and an open-source software stack for training new behaviors. Default actions include walking, recovery, and object picking. First deliveries were targeted before Christmas 2026; new orders are now estimated at four to six months. The video shows the robot’s size, gait, and recovery. Demand figures reported publicly exceed 10,000 units early and later moved higher. The robot is a maker and research platform rather than a finished consumer appliance. No large-scale customer reliability data have been published because shipments are still pending. The combination of price and open training tools has driven the response.

NEURA Robotics has introduced Neuron AI, a multimodal foundation model for Physical AI developed and trained in-house in Metzingen, Germany. The model uses a Perception-Language-Action architecture that combines 3D vision, tactile and force signals, audio, robot state, and language into a shared representation updated over time. A memory layer is included so the robot can retain context when objects leave the camera view. Neuron AI is trained on data from robots operating in NEURA Gyms, from simulation, and from people wearing sensor-equipped suits. The company states that the same model runs across five of its robot embodiments and their grippers and hands. Customers can fine-tune the model for their own processes through the Neuraverse platform and deploy the results to fleets. The video features CEO David Reger and Business Unit Head Dr. Jens Buchner explaining the approach. No independent benchmarks of task success rates or latency have been published. The announcement was made in mid-October.

Vision-language-action models primarily use camera images. Neuron AI adds touch, force, and audio so the robot can respond to events that cameras alone cannot capture, such as the click of a connector or the resistance of a screw. The memory component is intended to support object permanence and causality. Training data come from three sources: real robot interactions in NEURA Gyms, simulated variation, and human demonstrations. The company plans to expand the Gyms globally to increase data volume and reduce the time required to train new skills. The model is available for customer fine-tuning via Neuraverse. The video is an explanatory interview rather than a task demonstration. NEURA has previously shown its 4NE1 humanoid performing longer sequences. No quantitative comparison against vision-only models has been released. The claims of multimodal fusion and memory are company statements. Independent verification of performance gains has not been published.

NEURA Robotics builds several robot platforms, including the 4NE1 humanoid and industrial arms. Neuron AI is positioned as a shared foundation across these embodiments. The company says the model encodes the different sensor streams in one space over time and produces a single stream of action. Fine-tuning by customers is offered so that site-specific processes can be added without retraining the entire model from scratch. The video shows the executives describing the architecture and the training approach. It does not include side-by-side task results or failure-mode analysis. For operators evaluating foundation models for physical systems, the relevant questions are the data volume required for a new skill, the latency of the closed loop when touch and audio are included, and the safety margins during contact-rich tasks. Those details are not provided in the announcement or the video. The model is an architecture and training update rather than a finished, independently benchmarked product.

Physical AI models that incorporate touch and audio remain less common than vision-centric systems. NEURA’s approach treats the additional sensors as first-class inputs rather than optional add-ons. The memory layer is described as necessary for tasks in which objects move out of view. The company has stated that Neuron AI is the only multimodal foundation model of its type developed in-house in Germany. No third-party confirmation of that uniqueness claim has been published. The video is consistent with the written announcement. It provides an accessible explanation of the PLA architecture and the three training data sources. Operational results on customer sites have not yet been reported. The Neuraverse platform is the stated channel for fine-tuning and fleet deployment. The announcement is a model and platform update. Performance data remain company-reported.

NEURA Robotics has introduced Neuron AI, a multimodal foundation model that combines 3D vision, touch, audio, robot state, and language with a memory layer. The model is trained on robot data from NEURA Gyms, simulation, and human demonstrations and runs across five robot embodiments. Customers can fine-tune it via Neuraverse. The video features company executives explaining the architecture. No independent benchmarks have been published. The model is positioned as a shared brain for the company’s robots. Claims of multimodal advantage and reduced skill-training time are company statements. The announcement is an architecture and training update. Operational results on deployed fleets remain unreleased. The video is an explanatory overview of the model.

Finnish integrator Siipotec has delivered an automated production system for ET Listat that combines KUKA industrial robots, a 36-meter linear track, and autonomous mobile robots. The line sorts cut-to-length wood products and directs them to up to 18 stations based on length and quality. Two KUKA robots retrieve finished layers from sorting bins and perform handling operations that include 180-degree rotation around two axes. A third KUKA robot mounted on the 36-meter linear track assembles individual product layers into finished bundles. AMRs then transport the bundles. Siipotec describes the installation as an award-winning solution that improves productivity, consistency, ergonomics, and material flow while reducing waste. The video shows the robots, track, and AMR movement inside the ET Listat facility. No independent throughput measurements or before-and-after production figures have been published outside company statements.

Wood-processing lines often involve long, heavy, and variable-length pieces that are difficult to handle manually. The combination of fixed robots, a long linear track, and mobile robots allows the system to cover a large area without requiring workers to move between stations. The linear-track robot can reach multiple sorting and bundling positions. Servo-controlled guides and positioning systems are used to maintain accuracy. Siipotec specializes in automation for the mechanical wood industry and has previously delivered robotic cells and production lines for Finnish manufacturers. The ET Listat project is presented as a complete material-handling and internal-logistics solution. The company has not released cycle times, uptime percentages, or energy-consumption data. The video illustrates the physical layout and the movement of the robots and AMRs. It does not include quantitative performance metrics.

KUKA industrial robots are widely used for material handling, packaging, and assembly. Mounting one on a 36-meter track extends its working envelope for long products. The AMRs handle the final transport step, removing the need for fixed conveyors across the entire facility. Siipotec integrated the robots, track, vision or sensing systems, and AMR fleet control. The installation is one of several wood-industry projects the integrator has completed. No customer quote or specific award name is given in the publicly available description beyond the general “award-winning” characterization. The video is a customer-success overview rather than a technical deep dive. For manufacturers evaluating similar lines, the relevant questions are the actual throughput gain, the maintenance requirements of the long track, and the reliability of the AMR handoff. Those details are not provided in the promotional material.

ET Listat manufactures building mouldings, profiles, decoration panels, and wooden floors. The company produces large volumes of cut-to-length products that must be sorted, bundled, and moved. The Siipotec system automates the sorting, bundling, and internal transport steps. The 36-meter track robot and the AMR fleet are the two most visible elements in the video. The project demonstrates how industrial robot arms, linear motion, and mobile robots can be combined for variable-length material handling. No data on worker reassignment, injury reduction, or long-term reliability have been released. The video shows the system operating in the factory. Claims of higher productivity and improved ergonomics are company statements. The installation is a practical example of integrated industrial automation rather than a humanoid-robot application.

Siipotec has installed a KUKA-based automation system at ET Listat that uses industrial robots on a 36-meter linear track and AMRs to sort, bundle, and transport wood products. The solution is described as award-winning and is intended to improve productivity, quality, and ergonomics. The video shows the robots and mobile units in operation. No independent performance metrics have been published. The project is a conventional industrial-robot and AMR installation for the wood industry. It demonstrates integrated material handling across a large floor area. Quantitative results remain company-reported. The system is an example of production-line automation rather than humanoid deployment.

Luwu Dynamics has published the hardware, runtime, and training resources for XGO-Duck, a 3D-printable biped adapted from Pollen Robotics’ Microduck. The robot uses an Arduino UNO Q (Qualcomm processor running Linux plus STM32 microcontroller), 15 Feetech 1910 serial servos, a custom expansion board with a QMI8658A IMU, and printed structural parts. The supplied runtime includes walking, get-up, and pick policies that run on the UNO Q. The hardware repository contains STL files, PCB designs, a bill of materials, and an assembly guide. The project is explicitly credited to Pollen Robotics’ Microduck and microduck_rl work. The video introduces the hardware, assembly, on-robot software, and simulation workflow. No commercial production timeline or large-scale user base has been announced. The robot is a maker-oriented open-source platform rather than a finished consumer product.

Microduck is a small duck-shaped biped with 15 degrees of freedom that Pollen Robotics has shown walking and interacting. XGO-Duck retains the body shape, joint layout, and joint order but replaces the electronics and servos. The Arduino UNO Q splits compute: the Linux side runs policy inference and a web interface, while the STM32 handles real-time servo and sensor communication. The expansion board manages power, the servo bus, and the IMU. Luwu Dynamics has released ONNX policy files for walking, recovery, and picking so builders can start without training their own models. A separate training repository allows users to explore reinforcement learning. The video shows the assembly process and the robot executing learned motions. Independent long-duration reliability data from multiple builders have not been published. The project is an open hardware and software adaptation intended for makers who want to experiment with biped locomotion and on-device policies.

Open-source robot platforms lower the barrier for experimenting with learned motion policies. XGO-Duck provides printable parts, a documented electronics stack, and ready-to-run policies. Builders still need to source the Arduino UNO Q, the 15 servos, the expansion board, and batteries, then calibrate the joints. The runtime includes a web interface for control and monitoring. The training repository is separate and requires a suitable GPU machine for new policy development. The video walks through the hardware, assembly steps, and the transition from simulation to the physical robot. No formal safety certification or commercial support package has been announced. The project is suitable for research and hobby use. Claims of reliable walking or picking are limited to the supplied policies under the conditions shown. Broader performance depends on individual builds and calibration.

Pollen Robotics previously demonstrated Microduck and has made related software available. Luwu Dynamics adapted the design for the Arduino UNO Q and Feetech servos and published the full set of files. The expansion board is a custom design that integrates the IMU and servo interfaces. The video shows both the mechanical assembly and the software side, including policy execution. The project does not include a camera or advanced perception in the base configuration. Additional sensors can be added by builders. No large community deployment numbers have been reported. The repositories are public on GitHub. For makers and students, the combination of printable mechanics, dual-chip compute, and learned policies offers a complete starting point. The video is a project overview rather than a controlled benchmark.

Luwu Dynamics has released open-source hardware and software for XGO-Duck, a 3D-printable biped based on Pollen Robotics’ Microduck. The robot uses an Arduino UNO Q, 15 Feetech servos, and a custom expansion board. Supplied policies cover walking, get-up, and pick. The video shows hardware, assembly, and on-robot software. The project is a maker platform rather than a commercial product. No independent large-scale reliability data have been published. Builders can start with the supplied policies or explore the training repository. The adaptation is credited to the original Microduck work. The robot is intended for hands-on experimentation with learned motion.

NVIDIA CEO Jensen Huang delivered the GTC Berlin 2026 keynote on October 21. Announcements included the Agile ONE SN reference humanoid developed with Agile Robots, scheduled for early 2027; RTX Spark Windows laptops with up to 128 GB of memory for local AI agents; the Vera Rubin DSX Pod, a containerized AI factory design that NVIDIA says can deploy in as little as 14 days at a prepared site; and confirmation that Vera Rubin is in full production, with Huang claiming another roughly 30-times performance leap over Blackwell. OpenShell and Sentry security tools for autonomous AI agents were also introduced. The video is a third-party recap that also covers Unitree’s H2-X demonstration. All performance and availability figures are company claims. No independent benchmarks were released with the keynote.

The Agile ONE SN is a 31-DoF reference platform built on NVIDIA Isaac GR00T and Jetson Thor. NVIDIA says it ships preconfigured with compute and software to reduce researcher setup time from weeks to days. Agile Robots is targeting Q1 2027 availability. RTX Spark is a line of Windows PCs with RTX graphics and up to 1 petaflop of local AI performance. Vera Rubin NVL72 systems are liquid-cooled and connect 72 GPUs per rack. The DSX Pod packages these systems in modular containers for faster site deployment. OpenShell is an open-source runtime for isolating and monitoring AI agents. The video attributes every figure to the announcing company. Huang’s keynote also addressed European AI infrastructure and Hyperion vehicle platforms. The humanoid and compute announcements are the portions most relevant to robotics.

Reference platforms are intended to give developers a validated starting point rather than a finished commercial product. The Agile ONE SN is one such system. NVIDIA has not published pricing or production volumes. Vera Rubin’s “full production” status has been stated at multiple events; racks are reported to be running at customers. The 30-times claim is a company comparison under specific workloads. Independent measurements of end-to-end robot performance on the new stack have not been released. The video recap includes the Unitree H2-X high-impact demonstration alongside the GTC announcements. It does not add new technical data. For teams evaluating onboard compute or simulation tools, the Jetson Thor and Isaac GR00T updates are the practical takeaways. Availability timelines remain early 2027 for the reference humanoid and DSX Pods.

NVIDIA has expanded its robotics platform with GR00T models, Isaac tools, and Jetson Thor modules. Multiple companies, including Agile Robots, are building on the stack. The GTC Berlin keynote presented these elements together with data-center and personal-AI hardware. The video is a summary that also covers Unitree’s recent H2-X video. No new customer deployments or quantitative robot task success rates were announced. The reference humanoid is positioned for research and development. RTX Spark targets local agent workloads on Windows. Vera Rubin targets large-scale inference and training. All claims are attributed to NVIDIA or the partner companies. The keynote is a platform update rather than a product launch with immediate availability.

NVIDIA’s GTC Berlin keynote highlighted the Agile ONE SN reference humanoid, RTX Spark laptops, Vera Rubin in full production, and the DSX Pod containerized AI factory design. OpenShell agent security tools were also introduced. The video is a recap that includes Unitree’s H2-X demonstration. All specifications and timelines are company statements. The Agile ONE is scheduled for early 2027. No independent performance data were released. The announcements expand NVIDIA’s robotics and AI infrastructure offerings. Operational results on robots will depend on subsequent integrations. The keynote is a platform and hardware update.

XPENG has released video of its IRON humanoid adjusting mid-interaction during handshakes, fist bumps, and high-fives. When a palm closes into a fist or the high-five position changes, the robot switches hands or modifies its motion without completing a fixed routine first. The company attributes the behavior to an upgraded VLA 2.0 model with stronger temporal understanding and future-behavior prediction, running on-device. XPENG says the robot reassesses the scene approximately every 0.2 seconds and predicts action changes within about six seconds. The demonstration is a controlled interaction test. No independent latency measurements or success rates across varied users have been published. The video shows continuous observation and adjustment rather than a single pre-scripted sequence. IRON is XPENG’s general-purpose humanoid powered by the company’s Physical AI stack.

Real-time motion following requires the robot to update its plan while an action is already under way. Fixed gesture recognition followed by a complete motion is simpler but less flexible when a person changes intention. XPENG’s description of continuous observation and on-device VLA 2.0 inference is consistent with the visual result in the video. The model is also used in the company’s vehicles. Transferring the same foundation to the humanoid is a company claim. The 0.2-second reassessment interval and six-second prediction horizon are XPENG figures. No third-party timing data have been released. The video is useful as a qualitative demonstration of adaptive social interaction. It does not show the robot performing the same behavior in unstructured public settings or under varying lighting and crowd conditions.

IRON features a spine-like structure, bionic muscle layout, and soft skin. The interaction sequence tests upper-body coordination and the ability to switch limbs when the target moves. XPENG has previously shown IRON walking, tracking speakers, and holding multi-language conversations. The current video focuses on physical gesture following. The company notes that the footage is from development and testing and that the final product may differ. No force or contact-safety data for the handshake interactions have been published. For researchers studying human-robot interaction, the relevant questions are the latency of the closed loop, the robustness of the prediction model when intentions change rapidly, and the safety margins during contact. Those details are not included in the promotional clip. The demonstration illustrates on-device temporal reasoning applied to social motion.

VLA 2.0 is XPENG’s vision-language-action model. The company has deployed it in vehicles and is now applying an upgraded version to IRON. The on-device claim means inference occurs on the robot’s compute rather than in the cloud. IRON carries multiple Turing chips. The video does not show the internal compute configuration. Continuous adjustment during a handshake is a limited but visible test of the loop. Broader claims of general social competence would require longer, more varied trials. No such data have been released. The demonstration is consistent with XPENG’s positioning of IRON as a platform that combines physical motion with real-time AI. The video is the primary public evidence of the adaptive interaction behavior.

XPENG has shown its IRON humanoid adjusting in real time during handshakes, fist bumps, and high-fives. The company attributes the behavior to an upgraded on-device VLA 2.0 model with temporal understanding and prediction. The robot is said to reassess the scene every 0.2 seconds. The video demonstrates continuous observation and motion adjustment. No independent metrics have been published. The sequence is a controlled development demonstration. Final product behavior may differ. The clip illustrates adaptive social interaction on the IRON platform.

Unitree has released a promotional video titled “H2-X Kung Fu Legend Guardian.” The clip shows the humanoid performing high-impact martial arts movements, jumping from a platform labeled four meters high, landing, and recovering. It also depicts staged protection and rescue scenes, including intervening with a knife-wielding figure and lifting the front of a car so a dummy can be pulled free. Unitree states the footage is real and has not been sped up. The company lists a maximum endurance of approximately eight hours while standing. No detailed hardware specifications, pricing, or availability for the H2-X have been published. The video is a dynamic capability demonstration rather than a product announcement with technical data. The standard Unitree H2 is approximately 182 cm tall with 31 degrees of freedom; the H2-X relationship to that platform is not specified beyond the name.

High-impact demonstrations test joint strength, whole-body coordination, and impact recovery. The H2-X sequence includes kicks, punches that damage a car door, aerial flips, and a high drop onto mats. The protection and car-lift scenes are staged. Unitree has not published force measurements, success rates across repeated trials, or energy consumption during the stunts. The eight-hour standing endurance figure is a company claim. Dynamic operation under load would reduce that figure. The video has circulated widely and generated discussion about the physical power of current humanoid platforms. It does not address safety systems, autonomous decision-making, or real-world deployment constraints. The company includes a note to use the robot safely and in a friendly manner. The demonstration is visual evidence of dynamic motion capability.

Unitree has previously shown its robots performing parkour, flips, and high-speed locomotion. The H2-X video continues that pattern with higher-impact actions and a guardian narrative. The car-lift sequence requires significant torque at the arms and torso. Whether the motion is fully autonomous or partially teleoperated is not stated. No independent verification of the eight-hour endurance or the authenticity of every frame has been released. For researchers, the relevant questions are the control policies that generate the motions, the actuator limits that allow the impacts, and the recovery behaviors after falls. Those details are not included in the promotional clip. The video is useful as a qualitative illustration of the platform’s dynamic range. It is not a technical datasheet.

Unitree is a Chinese company known for commercial quadruped and humanoid platforms. The H2 series has been marketed for research and industrial use. The H2-X video adds a security-themed presentation. No contracts or deployments for a guardian role have been announced. The demonstration does not show the robot operating unsupervised in public spaces or interacting with real people outside staged scenes. Battery life, compute architecture, and sensor suite for the H2-X remain undisclosed. The video is the primary public evidence of the platform’s high-impact capabilities. Claims of real, unsped footage are company statements. Quantitative performance data have not been provided.

Unitree has released a video of the H2-X performing martial arts, high drops, and staged protection and rescue actions. The company states the footage is real and lists approximately eight hours of standing endurance. No detailed specifications or independent verification have been published. The video is a dynamic capability demonstration. The guardian narrative is staged. The clip shows the physical range of the platform under controlled conditions. Real-world autonomous security use has not been demonstrated. The demonstration is visual evidence of high-impact motion.

UK-based Humanoid has announced an expanded collaboration with NVIDIA to develop a high-performance robotic edge runtime for its next-generation robots. The companies are integrating NVIDIA Jetson Thor, NVIDIA Isaac OS, NVIDIA Holoscan Sensor Bridge, and the Holoscan SDK. Humanoid says the move to Jetson Thor has allowed it to consolidate two onboard computers into a single system while maintaining computing capacity, with potential additional headroom. Unified CPU and GPU memory on Jetson Thor is expected to reduce data movement and latency. Holoscan Sensor Bridge places camera data directly into GPU-accessible memory. Isaac OS provides a GPU-native runtime for coordinating perception, inference, and control. Multi-Gigabit Ethernet supplies up to 100 Gb of bandwidth for sensor and control signals. The company has used NVIDIA Isaac Lab and Isaac Sim since its early days for training and simulation. The announcement was made on October 21. The video is the company’s general platform overview.

Onboard compute is a constraint for humanoid robots that must process high-bandwidth sensor data and run AI models in real time. Moving from two computers to one reduces hardware complexity, power draw, and integration effort. Humanoid states that the new architecture simplifies how different workloads share resources. The collaboration does not include published benchmarks of inference latency, power consumption, or task success rates on the new stack. NVIDIA has released Jetson Thor modules and Isaac tools for general robotics development. Humanoid is one of several companies adopting the platform. The press release quotes CEO Artem Sokolov on the need for significant computing power and real-time integration of sensing, AI, and control. No deployment timeline or robot model name for the next generation is given. The work is framed as foundational architecture rather than a finished product release.

Humanoid positions itself as a UK AI and robotics company focused on commercially scalable and safe humanoid robots. Earlier work has included proofs of concept in industrial settings and the use of NVIDIA simulation tools. The expanded collaboration extends that relationship into the onboard runtime. Holoscan is intended to lower latency between sensors and models. Isaac OS is intended to coordinate multiple robotics workloads on the GPU. These are engineering choices rather than demonstrated performance gains. Independent measurements of the consolidated computer’s effect on robot behavior have not been released. The video shows the company’s robots and facilities but does not specifically demonstrate the new Jetson Thor configuration. The announcement is a technology partnership update. Operational results will depend on subsequent hardware and software integration.

NVIDIA has been expanding its robotics platform with Jetson Thor, Isaac GR00T, and related tools. Multiple European and global robotics companies are building on the stack. Humanoid’s announcement is consistent with that broader trend. The company says the technologies will make the platform easier to develop, test, and scale. No customer deployments or production volumes tied to the new compute architecture have been stated. The October 21 press release is the primary source. The video is relevant as company context but does not contain the technical details of the collaboration. For observers tracking compute choices in humanoid development, the consolidation to a single Jetson Thor system is the concrete engineering claim. Performance data remain unreleased.

Humanoid is expanding its collaboration with NVIDIA to build an onboard edge runtime using Jetson Thor, Isaac OS, Holoscan Sensor Bridge, and related tools. The company says Jetson Thor consolidates two computers into one while maintaining capacity. Unified memory and direct sensor-to-GPU paths are expected to reduce latency. The announcement was made on October 21. No independent benchmarks have been published. The video is a general company overview. The collaboration is an architecture and development partnership. Operational results have not yet been demonstrated. The work is intended to support the company’s next generation of robots.

UniX AI has released video of its Panther wheeled dual-arm robot performing facility patrol tasks. The robot is shown moving through low-light outdoor areas, office floors, and parking garages while using multimodal perception, adaptive manipulation, and anomaly detection. Panther is powered by the company’s UniTac, UniCortex, and UniFlex systems. UniX AI describes the platform as capable of 24/7 autonomous patrols in unstructured environments. The robot features mass-produced 8-DoF bionic arms, adaptive grippers, and an omnidirectional four-wheel steering and drive chassis. Earlier this year the company announced real-home deployments of Panther for household tasks. The facility-patrol video extends the same platform to security and inspection scenarios. No independent verification of continuous unsupervised operation has been published. The clip is a company demonstration of the robot in mixed indoor and outdoor settings.

Facility patrol requires navigation in spaces that change, detection of anomalies, and occasional physical interaction with equipment or barriers. UniX AI’s materials list fire-equipment inspection, route planning, and multi-floor coverage among the intended uses. The video shows the robot moving and performing inspection-style actions. It does not include long-duration logs, false-positive rates, or comparisons with human patrols. The company has previously claimed monthly production volumes above 100 units for its wheeled platforms and has positioned Panther for commercial service, home, and industrial roles. The patrol application is one of several the company markets. The video is relevant to that claim. Performance under real security contracts remains unstated in the public materials.

Wheeled dual-arm robots have advantages in indoor and mixed environments where stairs are limited. The omnidirectional chassis and vertical reach of the upper body allow Panther to access floor-level and elevated objects without repositioning the base for every task. UniX AI’s stack separates perception and manipulation (UniTac/UniTouch), long-horizon planning (UniCortex), and cross-scenario generalization (UniFlex). The company says these systems support the patrol behaviors. The video is the latest visual evidence. It does not replace quantitative testing of detection reliability or recovery from unexpected obstacles. For facility operators evaluating robotic patrol, the relevant questions are battery life, weather sealing, and integration with existing security systems. Those details are not fully specified in the current demonstration.

UniX AI is a Suzhou-based company founded in 2024. It has raised funding and begun deliveries of wheeled humanoid platforms. Panther is its third-generation full-size wheeled dual-arm robot. The facility-patrol video shows the robot in environments that include low light and mixed surfaces. The company presents the system as ready for 24/7 autonomous use. The evidence is the demonstration footage and the technical stack description. Independent long-term operational data have not been released. The video illustrates the platform’s mobility and inspection actions in the shown settings. Broader claims of continuous unsupervised security coverage remain company statements.

UniX AI has shown its Panther robot performing facility patrol in indoor and outdoor environments. The platform uses UniTac, UniCortex, and UniFlex and features 8-DoF arms on a wheeled chassis. The video demonstrates navigation, perception, and inspection-style tasks. No independent metrics of continuous autonomous operation have been published. The demonstration is consistent with the company’s security and inspection positioning. Real-world contract performance has not been detailed. The clip is a company demonstration of the patrol capability.

Mission Canvas is an open framework that lets developers define map-based, long-horizon robot tasks by combining LiDAR navigation, Behavior Trees, and Vision-Language-Action models. Users create a map, place waypoints, attach a Behavior Tree to each destination, and set the order of execution. Navigation moves the robot between waypoints. Behavior Trees organize the actions at each stop. VLA models supply adaptable manipulation where rule-based steps are insufficient. The framework is presented as a tool for experimentation and research rather than a finished commercial product. Code, documentation, and tutorial videos are available. The system builds on earlier Action Canvas work that defined single-location tasks. Mission Canvas extends that approach across multiple locations. The video demonstrates the workflow of mapping a space, designing missions, and watching the robot execute the sequence.

Long-horizon tasks require the robot to move, act, recover from local failures, and continue. Pure VLA policies can handle flexible manipulation but often lack reliable multi-location sequencing and safety checks. Pure rule-based systems are predictable but brittle when the environment changes. Mission Canvas places both under a Behavior Tree structure so that navigation, deterministic actions, and learned policies can be mixed at each waypoint. The developer can add new capabilities, such as pedestrian detection, by extending the tree. The video shows the interface and a robot carrying out a multi-location sequence. It does not include large-scale reliability statistics or comparisons against other orchestration tools. The value claimed is the ability to compose existing skills into repeatable multi-area missions without rewriting low-level code for every new workflow.

Behavior Trees are a standard way to structure robot tasks because they support sequencing, fallbacks, and parallel actions in a way that is easier to inspect than large neural policies. Adding VLA inference as a leaf node lets the tree call a learned policy when the situation requires adaptability and fall back to rule-based steps when precision or safety is required. Mission Canvas implements that pattern on top of LiDAR-based navigation. The open-source release includes the code needed to reproduce the workflow. Researchers can modify the trees, swap VLA models, or add new perception modules. The video is a demonstration of the current interface and execution loop. It does not claim that the framework solves general long-horizon autonomy. It shows one practical way to connect navigation, rules, and learned actions on a mapped space.

The framework is associated with ROBOTIS development tools and has been discussed in the Open Robotics community. Action Canvas handled single-workstation tasks. Mission Canvas adds the spatial layer. Both are positioned for developers who already have robot skills and want a way to deploy them across a building or workspace. The video walks through map creation, waypoint placement, Behavior Tree design, and execution. No production deployment numbers or customer case studies are attached to the tool. It remains a research and development framework. For teams building mobile manipulation systems, the open code and the clear separation between navigation, rules, and VLA are the practical contributions. The demonstration illustrates the intended workflow.

Mission Canvas is an open framework for map-based, long-horizon robot tasks. It combines LiDAR navigation, Behavior Trees, and VLA models so developers can define multi-location missions. Users place waypoints, attach task trees, and run the sequence. The video shows the mapping, design, and execution steps. Code and documentation are available. The tool is intended for research and experimentation. No large-scale reliability data have been published. The framework is a composition layer for existing robot capabilities rather than a complete autonomy solution.

Linkerbot has shown its L30 tendon-driven dexterous hand playing a rhythm game at high speed. The hand is a flagship model with a tendon-driven architecture that moves actuators out of the fingers, reducing distal mass. Published specifications list 17 active degrees of freedom plus passive joints, for a total around 21 joints, with fingertip force sensors, repeatability of ±0.20 mm, and joint speeds exceeding 400°/s on core motions. The hand weighs approximately 1.4 kg. Control is over CAN FD. The rhythm-game demonstration shows rapid, coordinated finger movements timed to on-screen prompts. Linkerbot has previously shown the L30 performing fine-motor tasks such as screw tightening and object manipulation. The game clip is a speed and coordination demonstration rather than a new product claim.

Tendon-driven hands can achieve higher finger speeds than designs that place motors in the fingers because the moving mass is lower. The L30 is marketed for research, industrial assembly, and tasks that require human-like dexterity. Company materials list fingertip forces in the 3.5–4 N range for individual fingers and higher grasp forces with multiple fingers. The rhythm-game video is one of several high-speed clips Linkerbot has released. It does not include independent timing measurements or a comparison against human players under controlled conditions. The visual result is a hand that keeps up with rapid note sequences. Whether the same control precision holds for longer, more variable tasks depends on the software and sensing used with the hand.

Dexterous hands remain a bottleneck for many humanoid and manipulation systems. A hand that can move quickly and return to precise positions is useful for assembly, tool use, and research into high-frequency control. Linkerbot’s L30 is one of several tendon-driven options available to labs and integrators. The rhythm-game demonstration is a clear visual of speed. It does not replace quantitative benchmarks on grasp success, force control, or durability under continuous industrial use. The hand is sold as a modular end effector that can be mounted on robot arms or humanoid platforms. The video shows the L30 in a controlled entertainment-style task.

Linkerbot is a Chinese company focused on dexterous hands. The L30 is positioned as its highest-performance tendon-driven model. Public specifications and demonstration videos are the main evidence available. The rhythm-game clip has circulated as an example of the hand’s speed. For researchers evaluating end effectors, the relevant questions are repeatability under load, thermal performance during continuous high-speed motion, and integration effort with existing controllers. Those details are not fully answered by a short game video. The demonstration is useful as a qualitative illustration of the platform’s response time.

Linkerbot’s L30 hand is shown playing a rhythm game at high speed. The hand uses a tendon-driven design with approximately 17 active degrees of freedom, fingertip sensing, and joint speeds above 400°/s. The video is a coordination and speed demonstration. The L30 is a commercial dexterous hand for research and industrial use. No independent performance metrics from the game sequence have been published. The clip illustrates the hand’s rapid finger control on a timed task.

NEURA Robotics has published further explanation of Neuron AI, the multimodal foundation model it introduced on October 14. The model uses a Perception-Language-Action architecture that combines 3D vision, tactile and force signals, audio, robot state, and language into a shared representation updated over time, with a memory layer for object permanence and causality. CEO David Reger and Business Unit Head Dr. Jens Buchner describe the approach in a new video. The company says vision alone is insufficient for many physical tasks; touch and sound confirm contact events that cameras miss. Neuron AI is trained on data from robots in NEURA Gyms, simulation, and human demonstrations in sensor-equipped suits. It runs across five NEURA robot embodiments and their grippers and hands. Customers can fine-tune skills via Neuraverse and deploy them to fleets. Inference is local. An 11-minute autonomous laundry sequence on the 4NE1 humanoid is scheduled for NVIDIA GTC Berlin.

The memory component is presented as necessary for physical action. A robot must remember that an object still exists after it leaves the camera frame and must track the consequences of its own prior actions. NEURA says the architecture was adapted specifically for these memory tasks. Training draws on three pillars: real robot data collected in the company’s gyms and environments, simulation that varies scenarios, and human motion captured with sensor suits. After base training across the five embodiments, the model is fine-tuned on a specific robot for individual skills. Time-to-Skill is the metric the company uses to measure how quickly a new task can be learned and performed reliably. No independent benchmarks comparing Neuron AI to other foundation models have been released. The laundry demonstration will be the first extended public test on hardware.

Multisensory fusion and memory address known limitations of vision-only models in unstructured settings. Tasks such as confirming a connector is seated, judging a weld seam by sound, or continuing a sequence after an object is occluded require information beyond the current camera frame. NEURA’s claim is that Neuron AI incorporates those signals by design. The evidence so far is the architectural description, the training sources, and the scheduled demonstration. The video with Reger and Buchner walks through the same points. Whether the model generalizes to customer environments and reduces time-to-skill as claimed will be tested when customers begin fine-tuning and when the laundry sequence runs in public. Until then, the model is a company system with a clear technical thesis.

NEURA has positioned Neuron AI as the intelligence layer across its robot family, including multiple generations of the 4NE1 humanoid and other bimanual platforms. The Neuraverse platform is the delivery mechanism for fine-tuned skills. The company has said it will scale NEURA Gyms globally to increase data and reduce training time for new skills. The October 14 announcement and the subsequent explanatory video are the public materials. The model is developed and trained in Metzingen, Germany. For potential customers, the open questions are the data requirements for fine-tuning, the reliability of transferred skills, and the cost of local inference on the robot. Those details are not fully specified in the current materials. The laundry task is the next scheduled public data point.

NEURA Robotics is providing additional detail on Neuron AI, its Perception-Language-Action foundation model. The model fuses 3D vision, touch, audio, robot state, and language with a memory layer, and runs across five embodiments. It is trained on robot, simulation, and human demonstration data. Customers fine-tune via Neuraverse. An 11-minute autonomous laundry sequence is planned for NVIDIA GTC Berlin. The new video explains the architecture and training approach. No independent performance numbers have been published. The model remains a company announcement with a scheduled hardware demonstration.

AheadForm is presenting its Elf-Xuan 2.0 humanoid at the World of Warcraft Infinite Party — Dalaran on Earth event in Shanghai. The company says it collaborated with World of Warcraft to give the robot a Skyborne-inspired appearance drawn from the new race in World of Warcraft: Forever. Elf-Xuan 2.0 features expressive facial movements, natural gaze, and coordinated body motion. The event runs October 16–18, 2026 at the National Exhibition and Convention Center (Hall 1.2H). AheadForm describes the robot as blurring the boundary between virtual characters and physical reality. The video shows the Skyborne-styled unit. The appearance is a promotional collaboration for the convention, not a claim of new general-purpose capabilities.

Elf-Xuan 2.0 is AheadForm’s lifelike bionic humanoid platform. The company has previously shown versions focused on facial expression and upper-body movement. Founder Yuhang Hu has described the work as bringing characters from games and virtual worlds into physical form for interaction and entertainment. The Shanghai event is a Blizzard and NetEase production marking 21 years of World of Warcraft in China, with a life-sized Dalaran recreation. AheadForm’s presence is a brand and character collaboration at that event. The robot is not described as performing autonomous tasks beyond the expressive demonstration. Specs, degrees of freedom, and price for Elf-Xuan 2.0 have been reported in secondary sources in broad ranges; the company materials for this event focus on the visual and interactive presentation.

Bionic humanoids designed for expressive interaction are a distinct category from industrial or warehouse platforms. AheadForm’s approach emphasises realistic faces, gaze, and motion so the robot can occupy entertainment and fan-event roles. The Skyborne collaboration places that capability inside a major game-convention setting. Attendees can meet the robot in person during the three-day event. The video is the company’s promotional clip for the appearance. No independent technical evaluation of the robot’s autonomy, reliability, or interaction quality at the event has been published. The demonstration is the physical embodiment of a virtual character for the duration of the show.

AheadForm is a Shanghai-based company focused on realistic humanoid heads and bodies for emotional expression and interaction. Its Elf-Xuan series has been shown at previous events. The World of Warcraft tie-in is a marketing collaboration that places the robot in a high-visibility fan setting. The company states that characters from games and films no longer have to exist only behind screens. The practical outcome is a styled humanoid present at the Shanghai convention. Whether the same platform moves into other commercial entertainment or interactive roles depends on further development and customer interest beyond this event.

AheadForm is displaying Elf-Xuan 2.0 with a Skyborne-inspired appearance at the World of Warcraft Infinite Party in Shanghai from October 16–18, 2026. The robot features expressive facial movements and coordinated motion. The collaboration is promotional for the event. The video shows the unit. It is a character embodiment demonstration rather than a new technical capability claim. The event is the public venue; the robot is the physical presence AheadForm is bringing to it.

Kinova has released a demonstration of two Gen3 robotic arms performing coordinated manipulation. The Gen3 is a lightweight, modular arm platform developed for research and professional applications. It supports grasping, vision-based manipulation, assembly, and robotics development work. Dual-arm setups allow researchers to explore coordinated movements and tasks that require two arms. The video shows the arms working together on a manipulation sequence. Kinova, based in Quebec, Canada, designs the Gen3 series for labs and industrial settings where portability, safety around people, and open software interfaces matter. The Gen3 is available in 6-DoF and 7-DoF configurations with options for integrated vision and various end effectors.

The Gen3 platform is widely used in research because of its ROS support, high-level and low-level control, and modular design. Dual-arm configurations appear in academic and industrial projects for tasks such as assembly, teleoperation, and collaborative manipulation. One documented use is an augmented-reality teleoperation system in which two Gen3 arms are controlled remotely inside a containment cell. The arms’ compact size and 7 degrees of freedom help with collision avoidance in constrained spaces. Kinova’s Kortex software and open APIs support custom development. The video is a company demonstration of dual-arm capability rather than a new product announcement. It illustrates coordinated motion that single-arm systems cannot perform without additional fixturing or sequencing.

Lightweight collaborative arms are common in research labs because they can be moved between experiments and operated near people with lower risk than large industrial robots. Dual-arm setups add the ability to hold one object while manipulating another, or to perform bimanual tasks such as folding, assembly, or tool use. The Gen3’s modular joints and optional vision modules make it suitable for vision-guided grasping and insertion tasks. Researchers use the platform with MoveIt, Gazebo, and custom controllers. The demonstration shows the arms executing a sequence; it does not claim autonomous high-level planning or long-horizon reliability. Performance depends on the software and sensing the user adds.

Kinova has positioned the Gen3 for education, research, and light industrial applications. Dual-arm work expands the range of experiments possible on the same hardware. Labs that already own single Gen3 arms can add a second unit for coordination studies. The video is a visual example of that configuration in action. No new specifications or software releases are announced. The interest for humanoid and manipulation research is the availability of a documented, commercially supported dual-arm setup that can be used for algorithm development and teleoperation studies. The arms themselves remain the same Gen3 platform; the demonstration is the coordinated use of two of them.

Kinova is showing two Gen3 robotic arms performing a coordinated manipulation task. The Gen3 is a modular research and professional arm used for grasping, vision-based work, and assembly studies. Dual-arm setups support exploration of bimanual tasks. The video is a company demonstration of the configuration. It does not introduce new hardware. Researchers use the platform for teleoperation, assembly, and algorithm development. The demonstration illustrates coordinated motion available on existing Gen3 hardware.

The Korea AI and Robot Industry Association announced on October 15, 2026 that Hyundai Motor, Qualcomm Technologies, and STATS ChipPAC Korea have joined the group. Hyundai Motor joins as a vice-chairman company. Qualcomm joins as a board member company. STATS ChipPAC Korea joins as a regular member. The association said it will use the companies’ expertise and networks to support policy development, proof-of-concept projects, international market expansion, and cross-industry collaboration. Standing vice chairman Bae Seok-tae said the new members strengthen the association’s ability to help South Korea lead in the global AI and robot market. The announcement comes ahead of RoboWorld 2026, scheduled for November 4–7 in South Korea.

Hyundai Motor has been a general member of the association since 2009. The elevation to vice-chairman status gives it a larger role in the association’s direction. Hyundai Motor Group has expanded its robotics footprint through the acquisition of Boston Dynamics and investments in AI and autonomous systems. Qualcomm’s participation brings semiconductor and edge-compute expertise relevant to robot platforms. STATS ChipPAC Korea is a semiconductor packaging firm; its membership adds advanced packaging capability to the association’s network. The association frames the additions as a way to break down barriers among automotive, semiconductor, AI, and robotics sectors.

The Korea AI and Robot Industry Association works on policy recommendations, field demonstrations, overseas market support, and cross-industry models. Adding large automotive and semiconductor companies is intended to broaden the industrial base available for those programs. Hyundai’s existing membership and the group’s broader robotics activities, including Boston Dynamics, give the association a direct link to a major Korean industrial player already investing in humanoids and physical AI. Qualcomm and STATS ChipPAC bring component and packaging perspectives that are relevant to the hardware side of robot development and production. The practical effect will depend on how the association uses the new members’ networks in specific projects.

South Korea has been expanding activity in physical AI and humanoid robotics, including regulatory sandbox pilots for factory deployments and investments by major groups. The association’s membership expansion is one institutional step in that broader activity. RoboWorld 2026 next month will be an early public venue where the new members’ involvement may become visible through exhibits, sessions, or joint announcements. The association’s website is korearobot.or.kr. The October 15 announcement is a membership update, not a product or technology launch. It signals continued consolidation of industrial participants around Korea’s AI and robotics policy and project framework.

Hyundai Motor, Qualcomm, and STATS ChipPAC Korea have joined the Korea AI and Robot Industry Association at different levels. Hyundai is a vice-chairman company, Qualcomm is a board member, and STATS ChipPAC Korea is a regular member. The association plans to draw on their expertise for policy work, demonstrations, and market expansion. The move comes ahead of RoboWorld 2026 in November. It is an industry-association membership announcement that reflects ongoing alignment between large Korean and international companies and the country’s AI and robotics sector. No specific joint projects were detailed in the announcement.

Google DeepMind researchers used Gemini 4 Argon in an internal test to watch a video of a human playing cornhole and generate the simulation and control needed for Apptronik’s Apollo 2 humanoid to sink seven consecutive shots on real hardware. The model mapped the human arm swing and release onto the robot’s joints, used the robot’s cameras to locate the board, built a MuJoCo physics simulation of the body and board, practiced the throw in simulation, and then transferred the result to the physical robot. On the real robot it adjusted release points based on where bags landed until seven shots went in. Google calls the workflow Video-to-Sim-to-Real and says it compresses weeks of manual modeling into a few hours. The demonstration was announced with Gemini 4 Argon on October 13.

Gemini 4 Argon is Google’s new frontier model, made available to developers and enterprises. In the cornhole test it was instructed to learn a physical skill outside its prior training distribution by watching ordinary video. The company reports that Argon scored 87 percent on its Embodied Reasoning benchmark, which tests understanding of physical and spatial problems in three-dimensional environments, and that it outperformed other frontier models on that test. The cornhole result is an internal demonstration of one throwing task. Google has not published details on board distance, total number of throws before the seven consecutive successes, or how the result generalizes to other objects, viewpoints, or contact dynamics. The video shows the sequence from human footage through simulation to the successful throws.

Apptronik’s Apollo is a humanoid platform that has been used in pilots with companies including Mercedes-Benz, GXO, and Jabil. The cornhole test is not a commercial deployment; it is a research demonstration of a training pipeline. The claimed advantage is the reduction in manual work required to build a usable simulation from video of a human performing a task. Once the simulation exists, the model can practice and then transfer to hardware, with further refinement from real-world feedback. Google has described similar pipelines on other tasks, including raising a robot’s success rate on a different manipulation task from 56 percent to 75 percent. Those figures are company-reported. Independent replication of the cornhole result has not been published.

Video-to-sim-to-real is one proposed way to expand the amount of training data available for physical AI. Ordinary videos of people performing tasks become potential sources of simulation and control targets. The practical limits are the accuracy of the inferred physics, the safety of transferring untested motions to hardware, and the need for real-world feedback when the simulation does not perfectly match the robot or the environment. The cornhole demo shows that, for this particular throwing task under controlled conditions, the pipeline produced a working result in hours rather than weeks of manual setup. It does not establish that the same approach will work for arbitrary industrial tasks or without human supervision during the real-world phase.

Google DeepMind has shown Gemini 4 Argon watching a video of cornhole, building a MuJoCo simulation, and enabling Apptronik’s Apollo 2 to sink seven consecutive shots on real hardware. The company says the Video-to-Sim-to-Real workflow reduces weeks of manual modeling to a few hours. The result is an internal demonstration of one skill. Argon’s 87 percent score on an embodied reasoning benchmark is a separate measure of the model’s performance on spatial and physical questions. The video documents the cornhole sequence; broader generalization and safety bounds remain open questions for future tests.

NEURA Robotics introduced Neuron AI on October 14, 2026. The company describes it as a multimodal foundation model for Physical AI and calls it the brain of the physical world. The architecture is Perception-Language-Action, or PLA. It combines 3D vision, touch and force data, audio, robot state, and language, then aligns them over time with a memory component so the robot can act on information that cameras alone do not capture and can retain context when objects leave view. The model is developed and trained in-house in Metzingen, Germany. Founder and CEO David Reger said an intelligent robot needs a brain that brings its senses together and remembers what happened. Independent benchmarks have not been published.

Neuron AI is trained on real-world robot data collected in NEURA’s own gyms, simulation-amplified data, and human demonstrations. The company says one model runs across five of its robot embodiments and their grippers and hands. Customers can fine-tune the model for their own tasks and deploy skills across fleets through the Neuraverse platform. Time-to-Skill is presented as the central measure of progress. Inference is described as local and real-time on the robot. A public demonstration is planned for NVIDIA GTC Berlin, where NEURA’s 4NE1 humanoid is scheduled to complete an eleven-minute laundry sequence fully autonomously. That sequence will be the first extended public test of the model on hardware.

The practical distinction NEURA draws is between models that primarily see the world through cameras and a model that also uses touch, sound, and memory of prior actions. Examples include confirming a lock has opened by its click and resistance, or gripping an object that is no longer in view. Object permanence—remembering that an object still exists after it leaves the camera frame—is presented as a requirement for physical tasks that cameras cannot solve alone. Whether those capabilities hold under real workplace variability is the open question. The company has not released quantitative comparisons. The architecture description and the planned laundry demo are the public evidence.

NEURA is a German robotics company that has positioned itself in the humanoid and cognitive robotics space. Its product family includes the 4NE1 humanoid across generations, wheeled variants, and other bimanual platforms. The Neuron AI launch is framed as the intelligence layer that can sit on top of those bodies and be fine-tuned by customers rather than rebuilt for each new task. For a buyer, the relevant questions are how long fine-tuning takes, what data the customer must supply, and how the model behaves when the task drifts from the training distribution. Those numbers are not in the announcement. The model is the offering; the laundry sequence is the scheduled public check.

Memory and multi-sensory fusion are harder problems than vision alone, and they are also the problems that limit many current robot systems in unstructured settings. A robot that can only act on what is currently in frame will fail when a tool is set down out of view or when contact must be judged by force rather than by sight. NEURA’s claim is that Neuron AI addresses those cases by design. The evidence offered so far is the architectural description, the training sources, and the upcoming demonstration. Until the laundry task or other independent tests are public, the model remains a company announcement with a clear technical thesis and a scheduled first long-horizon public run.

NEURA Robotics has released Neuron AI as its multimodal foundation model for physical robots. The model is designed to fuse 3D vision, touch, audio, robot state, and language with memory, and to run across multiple NEURA embodiments. Customers are expected to fine-tune skills and deploy them via Neuraverse. The first extended public test is an eleven-minute autonomous laundry sequence on the 4NE1 humanoid at NVIDIA GTC Berlin. The announcement is the model and the thesis; the demonstration will be the next data point. No independent performance numbers have been published yet.

Autowash is using KUKA industrial robots at its Colorado locations to perform touchless car washes. The system relies on KR CYBERTECH robots, 3D vision, and software that scans each vehicle and generates a customized path for washing and drying. Four cameras create a 3D model of the car; the planning software then produces a unique wash plan for that vehicle’s shape and attachments. The company says the approach delivers a faster, damage-free wash while reducing water and energy use. First commercial robotic bays are operating at the Lakeside location in Wheat Ridge. Autowash plans to roll the system out across more of its roughly 26–27 sites. The application is unusual for industrial arms that more commonly appear in factories.

Autowash was founded by Dennis and Erin Dreeszen and began with a single touchless location around 2013. The chain has grown to more than two dozen sites in Colorado. The founders have described the robotic system as a way to keep the wash completely touchless while improving consistency and resource use. In company materials they report washes in about four minutes versus a prior eight-minute cycle, water use around 20 gallons per car, and reductions in electricity compared with conventional touchless equipment. Those figures are company claims based on their own installations. Independent third-party measurements have not been published.

The technical approach is straightforward once the 3D scan exists. The robots carry custom tool heads with spray nozzles and air knives. Because the path is generated from the actual vehicle model rather than from a fixed template, the system can accommodate different body styles and attachments without a separate program for each. KUKA’s KR CYBERTECH series provides the reach and payload for the bay. The software layer handles the scan-to-path generation. The result is a wash that stays clear of the vehicle’s surfaces while still covering the body. That is the practical advantage over a fixed gantry or brush system designed around an average car.

Industrial robots are rarely the first choice for a consumer car wash. They cost more than conventional equipment, require a controlled bay, and need integration work that most wash operators do not perform in-house. Autowash’s decision to develop the system with KUKA and then deploy it commercially suggests the company sees a path to better margins or higher throughput once the capital is amortized. The chain already operates at a scale where a repeatable automated bay can be rolled out. The Lakeside site is the proof of concept in revenue service. Subsequent sites will test whether the same configuration can be replicated without custom engineering for every location.

The limits are the usual ones for any robotic cell that touches the physical world. A 3D scan can miss a trailer hitch or a damaged body panel. High-pressure water and air still require the nozzles to stay at a safe distance. Weather, customer behavior, and maintenance of the arms and vision system will all affect uptime. Autowash has not published failure rates or a full cost comparison against its previous equipment. What it has published is the architecture—scan, plan, execute with industrial arms—and the early commercial deployment. For the broader robotics industry the interest is the application itself: industrial robots performing a high-volume, consumer-facing service outside a factory.

Autowash is deploying KUKA KR CYBERTECH robots with 3D vision to generate per-vehicle wash and dry paths at its Colorado touchless car washes. The first commercial bays are running at the Lakeside site in Wheat Ridge, with plans to expand across the chain. The company reports faster cycles and lower water and energy use relative to its prior systems. Those are company figures from the installations. The system is an industrial-robot solution for a service that has historically used fixed machinery. The video shows the arms and the scan-to-path process; the business test is whether the same setup can be replicated profitably at the remaining sites.

DEEP Robotics has put a Lynx wheeled-legged quadruped on station at the Fuyang Grain Storage Area in China. The company says the robot is carrying out all-weather inspections and autonomous patrols. The machine combines wheels with four legs. On a flat yard it rolls. On steps or rough ground it uses the legs. That hybrid is the point of the design. A pure wheeled robot stalls at a curb. A pure quadruped is slower on a long flat patrol. The grain yard has both. The film shows the robot moving through the site. It is a deployment claim, not an independent audit of how many hours it has already run.

Lynx is part of a family. The industrial M20 weighs about 33 kg, carries a 15 kg payload, climbs steps up to 80 cm, and is rated IP66. It runs in temperatures from about -20 °C to 55 °C. The sport model is lighter, around 30 kg, with a lower payload and IP54. The company has used the same hybrid idea in other outdoor work: grape transport in Turpan heat, tea-leaf runs in Zhejiang, desert crossings. Grain storage is a different job. The value is a patrol that can go out in weather a person would rather skip, and that can handle a step or a rough path without a second machine.

Grain storage is a safety and maintenance problem as much as a logistics one. Temperature, moisture, pests, and access matter. Traditional wheeled robots need a prepared surface. Conventional quadrupeds can cross rough ground but give up speed on the long aisles of a yard. The wheeled-leg design is meant to keep both. DEEP Robotics has described the Fuyang work as all-weather inspection and autonomous patrol. Those are company claims. The film is the public evidence. A buyer who wants a shift length, a failure rate, or a cost per patrol will not find them in the clip. What the clip shows is a robot that can move through the site in more than one mode.

DEEP Robotics is a Hangzhou company focused on embodied AI and legged systems. The Lynx line is its answer to industrial outdoor work. The M20 was positioned as a mid-sized wheeled-legged robot for complex terrain and hazardous environments. Hot-swap batteries, SLAM, and dual-spectrum cameras for anomaly detection are part of the industrial pitch. Grain storage fits the pattern: a large outdoor site, weather, and a need for repeated patrols rather than a one-off demo. The company has put similar machines into vineyards, orchards, and deserts. A grain yard is the next setting on that list.

The limit is the same as any outdoor robot. A film of a machine moving through a yard does not say how long it runs between charges, how it handles a blocked path, or what happens when the weather is worse than the day of the shoot. DEEP Robotics has not published a shift log or a cost comparison with a human patrol for this site. The useful fact is narrower and real: a wheeled-legged robot is being used for inspection and patrol at a grain storage area, not only in a lab or a desert highlight. That is a practical application. It is not a proof that the category is finished.

DEEP Robotics has deployed a Lynx at the Fuyang Grain Storage Area for all-weather inspection and autonomous patrol, according to the company and the film. The hybrid design is the reason it can do both a long flat run and a step. The clip is the robot on the site. The numbers that would turn this into a buy decision are not in the video. The deployment itself is the news.

Officials from the Michigan Economic Development Corporation and the Michigan Department of Transportation visited Korean robotics companies Genix Robotics and T-Robotics on October 12–13, 2026. The Korea AI and Robot Industry Association was part of the meetings. The focus, according to the association, was business cooperation, technology exchange, investment opportunities, and expanding Korean robotics companies’ presence in North America, particularly in logistics automation and autonomous mobile robots. The visits are a continuation of Michigan’s longer effort to draw Korean manufacturing and technology investment into the state.

Michigan has been active in Korea. Governor Gretchen Whitmer led an investment mission to South Korea and Taiwan in September 2026, with the MEDC involved. That trip covered shipbuilding, semiconductors, robotics, drones, and future mobility. Korean companies already operating or expanding in Michigan include POEN, BRILS, HL Mechatronics, and Hansae Mobility, which announced a $93 million expansion and 102 jobs during the September trip. The October 12–13 visits to Genix and T-Robotics are a narrower robotics conversation inside that larger pattern. MEDC and MDOT on the same trip suggests both economic development and infrastructure or mobility interests are in the room.

T-Robotics has been building on its existing vacuum robot and AMR work. In late September it was selected as lead on a national project for AI-based mobile humanoid technology for display manufacturing, running through 2030 with government funding. Genix Robotics is a Korean robotics firm that has appeared in industry coverage around regulatory and market moves. The association’s account of the Michigan visits centers on logistics automation and AMRs, not on a signed deal. Meetings of this kind are how a company starts a conversation about a North American presence. They are not a plant announcement.

The practical interest for Michigan is logistics and manufacturing automation. The state has a dense auto supply chain, engineering talent, and a history of Korean investment. Korean robotics companies looking for a North American base have reasons to look there: customers, a manufacturing workforce, and an economic development agency that has already run missions and closed projects. The October visits put Genix and T-Robotics in front of MEDC and MDOT, with the Korea AI and Robot Industry Association facilitating. What comes next is a proposal, a pilot, or nothing. The meetings are the first public step.

The limit is that these are visits, not contracts. The association has described the focus as cooperation, technology exchange, investment, and North American expansion in logistics and AMRs. No investment figure, job count, or site has been announced from these two days. That is normal for an early meeting. The useful fact is that Michigan’s economic development and transportation agencies spent two days with two Korean robotics companies, alongside the industry association. It fits the pattern of the September mission and the existing Korean footprint in the state.

MEDC and MDOT visited Genix Robotics and T-Robotics on October 12–13, with the Korea AI and Robot Industry Association. The stated focus is cooperation, technology exchange, investment, and a larger Korean robotics presence in North America, especially logistics automation and AMRs. The meetings are the news. A deal, if one follows, would be a later story. Source: Korea AI and Robot Industry Association, korearobot.or.kr.

MindOn has put a number on the speed claim it made at the end of September. In an October 12 video, nearly an hour long, a dual-arm station folds shirts under Mind-1. The company says the run reached 142 garments, about one every 25 seconds, and labels the footage 1.0x speed and fully autonomous. A person throws the next shirt onto the table after each fold. The folding is what the company calls autonomous. The feed is a human. That is a company claim for a human-fed station, not a laundry that runs itself. The arithmetic is consistent with the number MindOn published. It does not independently confirm the video, the autonomy label, or any ranking.

Mind-1 is a model, not a body. MindOn introduced it on September 30 as physical AI aimed at human-speed task cycle times, and showed it on a dual-arm robot, a wheeled dual-arm, and a humanoid. Parcel handling, shirt folding, and a humanoid collecting laundry were the scenes, labeled the same way: real time, fully autonomous. The company said packaging and parcel sorting came in faster than a person in its own tests, and folding at 25 seconds against a human 15. Those figures are MindOn’s. They are cycle times on chosen tasks, not an independent time study, and they do not include retries or a person standing beside the table.

The hour-long film is the useful follow-up, because a two-minute demo cannot show whether the pace holds. A counter on screen climbs toward the company’s 142. The description talks about sustained operation, throughput, and reliability as the requirements for real work. MindOn has called the result the industry’s fastest robotic clothes folding. That ranking is a company claim. It would need a comparable test: same garments, same start, same fold quality, same help from a person. In the materials reviewed, that side-by-side comparison was not published. What the company has published is a long, labeled clip of one station, with a person feeding each shirt.

Speed is harder than it looks in a highlight reel. Slow teleoperation data teaches a slow policy, because the model learns the tempo of the demonstrations as well as the motions. At an end-effector speed of 3 m/s, 10 milliseconds of latency corresponds to about 3 cm of movement—enough to affect a precise grasp. Consecutive guesses that disagree make the arm correct itself, which at speed becomes a shake. MindOn says Mind-1 was trained on people working at natural speed, that inference latency dropped, and that a whole-body controller holds the trajectory. Those are design claims on the company blog. The film is the public check the company is offering: does the fold keep happening for an hour, at a counted pace, on the shirts someone keeps throwing onto the table.

A folding station is a narrow slice of laundry. Someone still has to bring the shirt, judge the fold, and move the pile. Other groups are working the handoffs around the fold, wash to dry to shelf, which is a different problem from cycle time at the table. MindOn’s number is the fold itself. A humanoid collecting laundry in a room, shown in the September film, is a different job from two arms at a green table. The model is meant to travel between those bodies. The October video tests it on one of them, for long enough to count, and leaves the rest of the workflow to the person standing next to the station.

Mind-1 is MindOn’s bid to make a robot finish work at a useful pace, not just finish it. The October 12 film is the company’s claim of 142 folds in an hour on a dual-arm station, with a person feeding each shirt. The autonomy label covers the fold, as MindOn frames it. The throughput is the company’s number. A shop that wants the number should ask what happens when the shirts change, when the fold quality is checked, and when nobody is standing at the table. Completing a task was the old bar. Doing it for an hour, at a counted pace, is the new one. This film is that count, and it is still a claim.

Put two Microducks in a room and they start to sing. The small biped from Pollen Robotics, the Bordeaux team that joined Hugging Face in April 2025, syncs with its neighbors over Bluetooth and joins a piece that is being made as it goes. The company’s clip says there is no score and no conductor, and that two, three, or four ducks are enough for a choir to break out. That is the product working as designed: an affordable robot with open-source software that does something visible the moment a second one walks in. Pollen’s published open-source offering covers its software and training stack. The product materials reviewed do not list complete hardware design files. The film is the choir.

Microduck is 25 centimeters tall and weighs under 800 grams. It has 15 motors, an articulated beak for grasping objects, a front camera, LiDAR, and two inertial measurement units. Those figures are from Pollen’s press kit. Pollen Robotics says it can walk, sit, crouch, roller-skate with the appropriate attachments, and recover from many common falls. Each unit generates a voice the first time it wakes and keeps that voice. It does not talk. It quacks and sings. Pre-order price is $399 before taxes and shipping, with first deliveries targeted before Christmas 2026. Seven trained moves ship in the box. The rest is for the owner to add.

The choir is the interesting bit of the software. Ducks find each other over Bluetooth, lock to a shared beat, and take parts. In the open code, the lowest duck takes the bass, a roster is shared, and the piece is chosen for that performance, then another after a breath. The marketing line is no score and no conductor. The mechanism is a beat beacon and a seating rule, which is how a small group stays in time without a person waving a hand. Mouth and head move with the part, so the choir is visible as well as audible. Two ducks are enough. A third joins what it finds. That is the demo, and it is also a small lesson in multi-robot timing.

This is Hugging Face’s second consumer robot, after Reachy Mini, the desktop companion that stays put. Microduck is the one that moves. Pollen’s point is physical AI you can put on a table and retrain: reinforcement learning in simulation, then on the duck, with the open software stack. Races, a kickabout, and a choir are the multi-robot demos, a way to try group behavior without a lab full of hardware. A $399 biped will not stock a shelf or fold a shirt. It will let a student see what a policy looks like when it falls and gets back up. That is a different product from a humanoid, and it should be read as one.

The limit is the size, and it is the point. A duck that sings with its neighbor is a demonstration of sync, identity, and a shared clock, packed into a body light enough to post. It is not a humanoid, and it should not be read as a step on the way to one. The open stack is the serious part: a public software runtime, a public training path, and a robot priced so a classroom can own two. The choir is how that stack shows itself without a slide. A buyer who wants a night aisle or a folding station is in the wrong room. A buyer who wants an affordable robot with open-source software is in the right one.

Microducks sing together because they can hear each other over Bluetooth and agree on a beat. Pollen and Hugging Face are selling that as a $399 robot with open-source software, first deliveries targeted before Christmas, with its own voice and a beak. The clip is the choir. The product is a small body you can retrain. There is no score on the film. There is a beacon in the code. Both are true. Open source, as published, covers the software and training stack. It is a duck that starts a song when another duck walks in.

CES 2027 is not a single hall that opens on a Wednesday. The week starts on Monday, January 4, with media day at Mandalay Bay and CES Unveiled, the preview where companies put a product in front of press two days before the floor opens to the industry. Tuesday, January 5, is the second media day. Conference programming starts that day, and C Space, the media and entertainment campus at ARIA, opens with it. The exhibit floor itself runs Wednesday, January 6, through Saturday, January 9, across the Las Vegas Convention Center, Venetian Expo, and a set of official venues that CES counts as three campuses: LVCC, the Venetian campus, and C Space. A badge that only thinks about the convention center is a badge that will miss half the week.

Hours on the floor are not the same every day, which is the detail that wrecks a late booking. Wednesday opens at 10 a.m. and closes at 6 p.m. Thursday and Friday run 9 a.m. to 6 p.m. Saturday closes at 4 p.m. A robotics team that lands on Saturday morning has already missed the keynote morning, the Unveiled briefings, and most of the meetings that get booked off the Wednesday aisle while people are still fresh. The useful map is the week, not the badge: Unveiled on the 4th, conference on the 5th, floor and stages from the 6th. The film on this page is a highlight from the last CES floor, robots beside other hardware. It is a picture of the mix. It is not the 2027 calendar.

The stages already named are mobility and computing, not a humanoid keynote, and that is worth knowing before anyone promises a robot moment. On January 6, the State of the Industry address is set for 8:30 a.m. in the Palazzo Ballroom at the Venetian. Stellantis chief executive Antonio Filosa follows, on connected vehicles and software across the group’s brands, with the talk tied to a West Hall exhibit so the audience can walk from the ballroom to the car. Waymo co-chief executive Tekedra Mawakana is listed the same day at 1 p.m. in that ballroom. Lenovo is taking its Tech World keynote to Sphere. None of these is a robot launch. They are the rooms the rest of the industry will be walking out of, which is where a humanoid meeting either gets booked or does not.

Foundry is the program closest to this beat. After a two-day debut in 2026, the AI and quantum destination returns to Fontainebleau for three days, January 6–8, with talks, demos, and the kind of hallway that software and hardware people actually use when they are not performing for a booth camera. Eureka Park remains the startup zone, the place a small humanoid team can still be found without a West Hall budget. CES Innovation Awards honorees are due in November, with embargoed products shown from January 4, which is another reason the media day matters more than the Saturday close. A short film of the last floor is a fair picture of how those zones feel in person. It does not tell you which room to be in at 8:30 on the 6th.

For a humanoid company, the week splits cleanly, and the split is the story. Unveiled is where a new hand or a new body gets a first write-up, before the floor noise starts. The exhibit floor is where a buyer walks the aisle and asks for a meeting that is not a stage demo with a marked path. Foundry is where the model people are, the ones who will ask what the policy was trained on. C Space is a different crowd, media and entertainment, useful for a bionic face and less useful for a payload chart. A team that only staffs a booth from Wednesday noon will meet the people who already had their briefing on Monday and have moved on. The last January showed how crowded the robot corner can get. The calendar above is how to be in the right room before that happens.

CES 2027 runs January 6–9 on the floor, with the real week opening January 4 at Mandalay Bay. Keynotes already set are Stellantis and Waymo on the 6th, plus Lenovo at Sphere. Foundry fills Fontainebleau for three days. Eureka Park is still the startup hall. The film is last year’s mix of robotics, hardware, and the Physical AI pitch. The program is the reason to book the earlier flight, and to know which campus you are actually walking into.

AheadForm brought Elf-Xuan 2.0 to Inclusion 2026 in Shanghai, and the point of the machine is the face. The company’s own clip, posted in September, holds on the eyes and the small movements around the mouth long enough that a viewer has to remind themselves it is a robot. AheadForm calls it a bionic humanoid, lifelike design and natural expression set on embodied AI. It is a showpiece, not a factory body, and the film is honest about that. Nobody in the frame is handing it a bin, a tool, or a shift sheet. The product is the moment a person across the room forgets, briefly, what they are looking at.

Inclusion, on the Bund, is an AI conference rather than a robotics fair, which is why a singing android fits the bill better than a palletizing demo would. Reports from the week said Elf-Xuan 2.0 performed a song written by its own system, with gesture and expression timed to the vocal, on a stage rather than in a booth corner. Coverage described synthetic skin and a face driven by many small actuators, the usual route to a microexpression that does not look like a mask sliding across a skull. Those are descriptions of a stage act, written by people who watched it. They are not a spec sheet, and AheadForm has not published degrees of freedom, payload, battery life, or runtime with the clip.

The company behind the face raised about $15.7 million in July 2026, in a round reported by the trade press rather than in a filing a buyer can read. That is enough to keep a small bionic team in parts, silicone, and stage time. It is not a humanoid factory round, and it should not be read as one. AheadForm’s public trail is the look: elves, singers, machines that hold a gaze and are filmed for the gaze. The Inclusion appearance is the clearest version of that bet, a body designed so the first reaction is recognition rather than a count of joints or a question about the hourly rate.

There is a real split in the humanoid market here, and the clip sits on one side of it without apology. Industrial machines are judged on a shift, a hand, a failure rate, and a price a plant manager can defend. Bionic machines are judged on whether a person across a table forgets, for a second, what they are looking at, and whether they stay to watch the next gesture. Both can be serious engineering. They are not the same product, and a buyer who confuses them will be disappointed by both. Elf-Xuan 2.0 is not applying for a night aisle in a Danish grocer or a line in a factory. It is applying for the room where people stay and watch.

The limit is the same as any stage robot, and it is the limit that keeps the rating at zero. A song, a look back, and a smooth turn of the head do not say what the machine does at 7 a.m., whether the face still tracks after an hour under conference lighting, or what it costs to keep the skin looking like that. AheadForm has not claimed a deployment, a wage, a lobby installation, or a second unit doing the same act the next night. The value of the September film is narrower and real: a Chinese bionic maker put a convincing face on a conference stage, and the audience treated it as a performer. That is the product working as designed.

Elf-Xuan 2.0 is AheadForm’s answer to the uncanny valley, shown at Inclusion 2026 in Shanghai and filmed so the eyes do the work. The clip is the company letting the face do the talking, which is a choice, not an omission. It is a bionic showpiece, funded like a startup, not an industrial humanoid, and it should be read as one. The aisle in Denmark and the line in a factory are different businesses. This one is the face.

DEEP Robotics has been running its DR02 humanoid on the streets of Hangzhou’s Xihu, or West Lake, district. In early October the company put a batch of the robots at busy intersections and at Xixi National Wetland Park, in visitor crowds and in sustained rain. The public clip shows navigation, obstacle avoidance, spoken reminders, and exchanges with people passing the corner. DEEP Robotics calls it a test of support for traffic work and tourist guidance. It has not called it a unit already on the city’s roster, and the film should be read that way.

At Zheda Road and West Wenyi Road in Lingyin Subdistrict, the robots stood at crossings in the wet. The company says they can spot common problems — a rider without a helmet, an overloaded e-bike, a pedestrian over the line, a vehicle past the stop bar — and answer with a voice prompt or a gesture rather than a ticket. In the rain clip, DR02 asks a rider to move back behind the line, calls a green light, and reminds people about helmets. IP66 protection on the whole machine is the hardware claim tied to working in that weather. No shift length, and no failure rate, was published with the film.

The other half of the trial is the scenic area, which is a different job from the intersection. West Lake and Xixi draw heavy domestic and foreign traffic in early October, and the repetitive questions — which path, which bus, what is closed today — eat the time of staff who are also meant to handle the thing that goes wrong. DEEP Robotics says DR02 took those questions in more than one language, and gave safety reminders along with route and transport answers. That is a service-window test. It is also the part a humanoid is easiest to film: a robot talking to a tourist, rather than a robot holding a post for a full morning beside moving traffic.

What the company has not said is as useful as what it has. It described a batch, not a number. It did not say how long each robot ran, whether a person was driving it from off camera, or how often the navigation held without a reset when a scooter cut the corner. A press note carried on the wires is still the company’s account of its own trial. Coverage since has treated the work as a street demonstration of a claimed all-weather design, and has been careful not to promote a test into a deployment the city has not announced.

The step from the lab to a wet intersection is still the right one to watch, which is why the clip earns the slot. A humanoid that can stand in rain, see a helmet missing, and say so in the pedestrian’s language is doing a job city staff already do, on a corner they already stand on. Whether it can do that job for a morning, beside moving traffic, without a handler in a van around the block, is the next question, and it is not answered by a one-minute film. DEEP Robotics says it will keep tuning DR02 for public-service work from the data these days produced. No timetable came with that line.

DR02 spent early October on Hangzhou roads and in a crowded park, in the rain, talking to people and pointing at traffic. DEEP Robotics is testing a helper for officers and for guides at the scenic edge of the same district. The film is that test, shot in the places the release names. A shift on the city’s payroll is not what was announced, and it should not be read into the clip.

Boston Dynamics has named Rohit Prasad chief executive, effective October 7, 2026. Prasad was senior vice president and head scientist for Alexa and artificial general intelligence at Amazon, the person associated with taking that assistant from a research bet to a product used at household scale. Hyundai Motor Group, which owns the company, said it wanted someone who had turned AI work into products people actually ship, not a research program that stayed in the lab. He follows Robert Playter, who left in February, and interim chief Amanda McMaster, who had been holding the seat.

The job in front of him is commercial Atlas, and the company has been making that case in public all year. Boston Dynamics argues that it already knows how to sell a robot: Spot and Stretch opened markets other firms are still explaining, and Atlas is the same play with a humanoid body. Learning in real environments, a design that can be built in volume rather than assembled as a prototype, and what the company calls physical intelligence are the three lines in its own film. Prasad’s hire is Hyundai putting an AI operator on top of a case the robotics team was already making.

The factory math is already public, and it is Hyundai’s number rather than a startup’s hope. In January, Hyundai said it aimed to build a plant able to make 30,000 robots a year by 2028, and to start putting them across its own sites that year. Later notes put Atlas at the Georgia metaplant from 2028, beginning with welding and logistics and moving toward component assembly by 2030. Boston Dynamics has also said the 2026 Atlas run is committed, including to Hyundai’s Robotics Metaplant Application Center and to Google DeepMind, with other customers planned for early 2027. That is a full year, spoken for, before the new plant exists.

The hand arrived just before the CEO, and it belongs in the same paragraph as the appointment. On October 1, Boston Dynamics showed a new electric-Atlas hand: four fingers, 13 degrees of freedom, no pinky, with encapsulated direct-drive actuators meant to be repeated on a line. The company tied it to mass manufacturing and to sim-to-real learning, not to a lab prototype that cannot be built twice. A humanoid sold into a car plant lives or dies on that hand, on whether it can take the tool the line already uses. Prasad inherits a body, a hand designed to be made, and a parent company that has already named the volume.

None of this is a shipped fleet, and the piece should not slide from a target to a delivery. Hyundai’s 30,000 is a factory target for 2028, and the site of that U.S. plant was still to be announced when the CEO news landed. Spot and Stretch are the commercial record, the robots already in customer hands. Atlas is the bet that the same company can do it again with a biped, under an executive who scaled Alexa from a demo to a business. The October 7 appointment is the organizational half of that bet. The hand shown six days earlier is the mechanical half.

Boston Dynamics has a new chief executive from Amazon’s AI ranks, a parent planning 30,000 robots a year by 2028, and a four-finger Atlas hand built to be made, not just shown on a stage. The film on this page is the company’s own argument for why that combination should win. The race Prasad was hired to run is the one from demonstration to a line Hyundai can actually fill.

The humanoid supply chain does not start with the face. It starts with the harmonic reducer, the roller screw, and the motor packed into a joint small enough to hide in a limb. At Guohua Intelligent in Qingdao, that path runs from those parts to integrated joints and then to bodies, some finished, some still open to the wiring. Reach farther with the same load and the shoulder wants more torque. Run it harder and heat becomes the limit. The finished walk only looks easy.

Guohua was founded in Qingdao in 2021 by Wang Guang to make those parts. The catalog runs from harmonic reducers and planetary roller screws to rotary and linear modules, then to arms and complete platforms. The company says it has more than a hundred reducer models. After a 2025 round that included Xiaomi’s manufacturing fund, it described a new base sized for 100,000 reducers, 20,000 screws, 5,000 joint modules, and 1,000 bodies a year. An earlier park plan used a higher joint number. Either way, the bet is volume in the part everyone else is short of.

High-end reducers were an import business for a long time. Chinese humanoid programs still buy a large share of gearboxes, screws, and encoders from a short list, and a late gearbox stops a body line even when the software is ready. Guohua’s case is that designing the reducer, the screw, and the joint together cuts that wait. A developer can take a module, an arm, or a whole platform. Payload, reach, and duty cycle decide which one. A heavier hand is a different joint, not a setting in the controller.

On the Qingdao floor the next layer is control. A command goes out over CAN or EtherCAT, an encoder reports where the joint actually is, and the drive corrects the gap. Temperature, faults, and a mechanical zero sit in the same loop. Push the response too hard and the arm vibrates. Guohua has also shown wheeled and two-legged platforms next to its arms, which is a reminder that the joint has to fit the job. A public-hall robot and a factory arm do not want the same stiffness, or the same cooling.

Hands still get the demos. Reducers and screws get the purchase orders. A supplier that can ship both the gearbox and the integrated joint shortens the path from a drawing to a tested arm, which is what the Qingdao visit is for. Guohua’s reported share of the domestic high-end reducer market is still small, on the order of a few percent, higher inside Shandong. That is a foothold in a business long held by overseas names, not a finished substitution. The visit does not open the order book.

For anyone building a humanoid this year, the scarce item is still the actuator. Guohua is one of the Qingdao shops trying to make that item at home, from the reducer up to the body. The useful question on a tour like this is not how many robots are standing in the aisle. It is whether the joint on the bench can be bought again next quarter, at the torque the arm actually needs.

Lunes Robotics does not want to win the body. The Augsburg company builds the software that is supposed to turn whatever humanoid a factory already has into a worker for ordinary industrial jobs. The line on its site is blunt: one platform, any humanoid, runs your night shift. Material transport, logistics, pick and place, and machine tending are the tasks it names. Navigation, locomotion, learning, and fleet orchestration sit in one stack, meant for cobots as well as humanoids.

That stack is called LUNAR. Under it, ARCS writes a robot program at design time and runs ordinary code on the shift, calling on AI only where a fixed script runs out. Lunes argues that a plant will accept a logged, inspectable action long before it accepts a model that cannot explain why the arm moved. Fleet tools are drawn to talk to the systems factories already run, MES, ERP, and WMS. The safety pitch is European certification from the start. A night shift is the promise. A named, certified line is not yet the public record.

A September clip shows the unglamorous half of that promise: recognize the object, then do the task. Lunes titles it object recognition of an Agibot G2 running LUNAR, and describes precision and flexibility working together. A lid and a sheet of paper are not the same grasp. Earlier, at SPS 2025, the same group showed machine tending on a CNC: open the door, take the finished part, put a blank in. Vision sorted the part. Demonstrated grasps taught the motion. The new clip is that idea on a humanoid, one brand of body, not a parade of them.

The people come out of factory automation rather than a research lab. Lunes cites more than 130,000 automation hours, a decade in industry, and more than twenty robotics and control engineers. Enes Ciritci is chief executive on the current site. A December 2025 interview introduced Jeff Hutz as founder of the Augsburg automation business and of the robotics spin-up. The names have not been reconciled in public. The work either way is the same: take commissioning experience from classic cells and aim it at a humanoid that has to live on an existing line.

“Any humanoid” is the sentence that has to earn itself. Task code that survives a new arm, a new hand, and a new camera is a different product from a demo on one G2. Lunes says it will start with narrow jobs so the first week has a number attached: parts moved, doors closed, blanks loaded. That is the right scope for a software house. It is also the scope a buyer should hold them to. A recognition clip can show the hand and the eye working together. It cannot show the 2 a.m. shift, or the second vendor’s robot.

Lunes is betting that the scarce piece in industrial humanoids is no longer the body but the shift software around it. LUNAR and ARCS are the names on that bet. The September film is a clean look at recognition tied to a task on an Agibot. The company still has to show the same job on another machine, on a real night, with the log a plant manager can read.

Linkerbot has posted a short clip of its LinkerHand units assembling a robotic arm. The video, published on 5 October 2026, carries the company’s own line: training on this task began last year, it has since moved onto Linkerbot’s production line, and it now forms part of how the arms are built. The footage is a process shot set to music, with no independent narration. That production claim is not confirmed outside the video and should stay attributed to Linkerbot.

The clip sits inside a longer factory story. Linkerbot, founded in 2023, builds dexterous hands in-house across linkage, direct-drive, and tendon designs, and has said it already uses its own hands on assembly lines to make more hands. In July 2026 it showed an upcoming O30 model it said could assemble robotic hands without a human in the loop. A Beijing line has been described as targeting tens of thousands of hands a year, alongside other Chinese factories. Those are prior company and trade reports, not a reading of the October clip.

What the 5 October video adds is a specific task: hands working an arm, not only hands building hands. Linkerbot says the skill left last year’s training runs and is now one of the steps used to build arms. The page does not show cycle time, a reject rate, or a count of arms finished this way. A trained grasp moved onto a line is a different claim from a station that runs every shift. Only the first is on screen.

The useful limit is the source. Linkerbot’s description is the only record that training began last year and that the arm step is now part of production. No outside plant visit is attached to this brief. Self-manufacture of hands is the better-documented thread: in-house parts, three drive routes, and a stated push toward automated assembly. The arm clip is consistent with that thread. It does not prove the arm task is routine.

Hands that build the next machine are the point Linkerbot has been making in public since the summer. The October short is a narrower exhibit of the same idea, aimed at an arm rather than a hand. Until a line log or a second source names the station, the sentence to publish is the company’s, not a confirmed process change. Price, throughput, and a customer for the finished arms are not in the clip.

Linkerbot’s 5 October video shows LinkerHand units assembling a robotic arm. The company says training began last year and that the task is now on its own line. That line claim is unverified outside the clip. The longer record is in-house hands and a Beijing capacity push, not this one station.

Xynova showed its Flex 2 hybrid-driven dexterous hand and Prima 1 direct-drive dexterous hand at IROS 2026 in Pittsburgh, which ran from 27 September to 1 October. In the company’s 30 September video, the two hands are described as a debut that drew experts, scholars, and representatives from institutions and companies to stop, try the hardware, and compare notes. That crowd line is Xynova’s. The clip is a booth short, not an independent count of visitors.

Flex 2 is the hybrid tendon-and-direct-drive design, about 400 grams with 23 degrees of freedom, built for reliable deployed manipulation. Prima 1 is the fully direct-drive hand, with 22 active degrees of freedom, built for force transparency, data collection, and algorithm training. Xynova’s pitch is that the pair splits those jobs: one hand for real-world operation, the other for the experiments that feed training. Both figures are the company’s published specs, also listed in a show briefing of IROS hand launches.

Pittsburgh was a United States debut, not a first public launch. Flex 2 was unveiled at ICRA in June 2026, and Prima 1 at the World Robot Conference in Beijing in August. Calling the pair major highlights of IROS is Xynova’s wording. Dexterous hands were a crowded category at the show, with many other exhibitors, so that ranking is not an independent judgment. The video does not establish a technical result beyond the booth appearance.

The company behind the hands is young and heavily funded relative to its age. Founded in 2024 and based in Hangzhou, Xynova is led by Alex Xia. It has raised about CNY 1.5 billion from investors including CATL, Xiaomi, JD.com, and Meituan, and says it is targeting annual capacity of 10,000 dexterous hands by the end of 2026. A Hangzhou production base has been described as already running. Those are company and trade figures, not a reading of the 52-second IROS clip.

The split between the two hands is the part worth keeping. Flex 2 is aimed at sustained work. Prima 1 is aimed at the sensing and control experiments that produce training data. A US debut of both, after Vienna, Shanghai, and Beijing showings earlier in 2026, is a distribution note. It is not evidence that either hand led the conference, and it is not a shipment announcement. No price is attached to this brief.

Xynova’s Flex 2 and Prima 1 appeared at IROS 2026 in Pittsburgh as a US debut. Flex 2 is the hybrid hand, about 400 grams and 23 degrees of freedom. Prima 1 is direct drive, with 22 active degrees of freedom. “Major highlight” is the company’s line. The world debuts were earlier, at ICRA and WRC.

The Robotics and AI Institute’s Roadrunner is a bipedal wheeled prototype that lifts one wheel, holds balance, then sets both wheels down and spins. RAI introduced the machine on March 23 as a research platform, not a product. It weighs about 15 kilograms, or 33 pounds, and has no torso. The legs are symmetric, so the knees can point forward or backward. The institute says that articulation is for obstacle avoidance and for changing how the machine moves through a space.

Roadrunner has three locomotion modes. Wheels sit side by side, or shift in line with one wheel behind the other, or the legs step. RAI said a single control policy covers both driving arrangements. Standing up from more than one position on the ground, and balancing on one wheel, transferred to the hardware with no extra tuning. Stair work, by rolling and by stepping, is part of the same demonstration. The one-wheel hold is the exposed pose: there are no arms to catch a tip.

RAI, formerly the AI Institute, built Roadrunner under its work on athletic intelligence and navigation in hard environments. The result it is claiming is zero-shot transfer of those behaviors, not a unique mechanism. Other wheeled-leg machines exist. RAI has not published a price, a customer, a battery figure, or a claim that the one-wheel balance runs off a prepared floor. A spin back onto two wheels shows recovery. It is not a payload test.

The useful distinction is the policy, not the trick. One controller for side-by-side and in-line driving, plus stand-up and one-wheel balance deployed without a second tuning pass, is what the March note actually says. That is a locomotion result on a 15-kilogram research body. It does not establish uptime, a payload, or a site where the robot is doing work. No order page has been attached to the prototype.

Two wheeled legs that can reverse at the knee are what make the sequence possible. The machine holds itself on one wheel, puts the second wheel down, and rotates. RAI has not published a product plan. The record is the balance, the mode switch, and the institute’s statement that several of those behaviors reached the hardware zero-shot.

RAI’s Roadrunner, about 15 kilograms, balances on one wheel and spins on two, and can drive side by side, in line, or step. The institute said a single policy covers both wheel arrangements and that stand-up and one-wheel balance transferred without extra tuning. It remains a research prototype, with no price and no named deployment.

Menlo Research has published the training and evaluation code for Asimov 1’s walking policies, including an Isaac Lab framework it calls Cyclotron. The September release covers reinforcement-learning workflows, checkpoint evaluation, and policy export, plus reward definitions and actuator settings tuned to this robot. A lab can train in simulation before it has a machine. Ordering a robot is separate from cloning the repository. The drop is software, not a new humanoid.

Asimov 1 is Menlo’s open humanoid for developers. The company lists 25 powered joints, a height of about 1.2 meters, and a mass of about 35 kilograms: six joints a leg, five an arm, one waist joint, and two in the neck. Motion control is on a Radxa CM5. A Raspberry Pi 5 handles media and networking. Sensing is an IMU, a monocular camera, a stereo microphone array, and a speaker. Power is a 13S4P lithium-ion pack. Rated arm payload is 5 kilograms, with a peak Menlo puts at 15. Those figures are the company’s.

On August 27, Menlo reported zero-shot transfer of a walking policy from simulation to the physical robot, with no per-unit tuning, on flat ground and on small variations such as slight slopes and grass texture. It said the same result held across more than one unit, and that the base policy runs onboard at 50 hertz on the motion board, driving 25 motors with no external computer in the loop. Training also moved between MuJoCo and Isaac Lab. That report is Menlo’s. An outside lab has not reproduced it in this record.

A public checkpoint is not a universal walker. Actuator settings are tuned to Asimov’s transmissions. A different motor or a different reduction may not stand up on the same reward. Menlo has said fleet-scale consistency and loco-manipulation are still ahead. MuJoCo and URDF models ship with the robot files. Full-body teleoperation and a client SDK are listed as in development. The open piece today is locomotion training and evaluation, plus the mechanical and electrical documentation.

The release sits with Viam only as developer infrastructure, and with Roadrunner only as locomotion research. Asimov is one body, repairable by design in Menlo’s wording, with a walking stack now in a public repository. No customer count and no commercial fleet are in the note. Wrist, head, and back panel are described as customizable. That is a research platform, not a shipped worker.

Menlo has opened Asimov 1’s locomotion training code on Isaac Lab, with the robot listed at 25 degrees of freedom, 1.2 meters, and 35 kilograms. The August zero-shot result, onboard at 50 hertz, is the company’s, limited to flat ground and small terrain changes. An outside reproduction is not in this report.

WUJI has mounted its dexterous hands on Gento Luna and shown the robot handling bottles, stacking blocks, and putting toys away. The September 26 demonstration is the company’s case that the fingers can change grasp with the shape and size of the object. WUJI is inviting other robotics teams to try the hands on further tasks, and pointed labs to its sales desk. No customer deployment, no price for the combined system, and no success rate have been announced with the demonstration.

Gento Luna is a wheeled, folding robot from Gento Robotics, offered for home service, research, inspection, and commercial work. The maker lists 26 degrees of freedom excluding the hands, an omnidirectional chassis, and force-controlled arms rated at 6 kilograms, with repeatability it puts near a few hundredths of a millimeter. Those are body specifications. They are not a measured lift from the bottle-and-block run, and WUJI did not publish a timed trial or a failure count.

WUJI builds direct-drive hands, with actuators in the fingers rather than tendons running to a forearm. The first public hand, shown in September 2025, has 20 active degrees of freedom across five fingers, a mass under 600 grams, and a stated fingertip force of about 15 newtons. Hand 2, teased in June 2026, kept that count and lowered back-drive torque to about 0.05 newton-meters. The Gento demonstration does not say which generation is on the wrists.

A bottle, a block, and a toy are three different grasps: a wrap, a pinch or cradle, and a light close on an irregular shape. WUJI presented them as everyday manipulation, not as a household shift. How often a grasp failed, and whether a person reset the scene between takes, is not in the release. The commercial offer is the invitation to other labs. A wheeled body with folding legs is the platform; the hands are the product being shown.

The demonstration sits apart from UBTECH’s plant and from Boston Dynamics’ commercial argument. It is a hand on a wheeled body, doing three short tasks in a controlled setting. No endurance hour and no named buyer are attached. Humanoid Press is treating it as a vendor demonstration of grasp, not as evidence that Luna is ready for an unsupervised home or a packing line.

Gento Luna, fitted with WUJI hands, handled bottles, blocks, and toys in a September demonstration. Grasp changed with object shape, which is the claim WUJI is taking to other teams. The body ratings are Gento’s. The hand generation on the wrists was not named, and no third party has timed the run or counted misses.

Boston Dynamics says it will win the humanoid race because it has already commercialized other robots. In a September 30 film made with Hyundai Motor Group and Bloomberg Media Studios, the company pointed to Spot and Stretch: markets it says it created, fleets it says it supports, and a payback it targets at under two years. Atlas, it said, is being built the same way, learning in real environments and designed to be manufactured, rather than remaining a research platform.

The figures in that case are about robots already in the field. Boston Dynamics said Spot is with more than 500 customers in 46 countries, and that it created the Stretch market and now leads it. Spot meets the two-year payback target, the company said. Atlas is aimed at the same bar. A humanoid market of $200 billion by 2035 was cited as a projection, not as orders in hand. The company did not publish an Atlas price or a delivery date for a paying customer.

Hyundai is the industrial partner. The companies opened a Robotics Metaplant Application Center in Georgia in September to train Atlas on manufacturing tasks, ahead of a rollout Hyundai has planned for 2028. Hyundai has also spoken of U.S. production capacity on the order of 30,000 units a year from 2028, with a large share it says is already reserved. Those are plans. Atlas is not yet a fleet on the scale of Spot, and a reservation is not a robot on a line.

The argument landed the day before Boston Dynamics unveiled a four-finger, 13-degree hand for Atlas, directly actuated and built for tools rather than a five-finger replica. Together, the two announcements are a commercial pitch: a gripper meant to be manufactured, and a claim that the company already knows how to keep robots working at customer sites. Winning the race remains a forecast. The hand is a product note. The race is not over.

Spot and Stretch are the part of the case that can be checked against a customer list the company has stated. Atlas at that scale is the goal. No shipment log, price, or unattended factory shift for the humanoid was part of the September 30 presentation. Humanoid Press is reporting the argument as Boston Dynamics made it, with Hyundai’s factory path attached, and not as a result.

Boston Dynamics is selling Atlas on the record of Spot and Stretch, plus a Hyundai factory path aimed at 2028. The 500-customer, 46-country line is the company’s figure for Spot. Atlas at that scale is still the aim, and the two-year payback is a target the company says Spot already meets and Atlas has not yet been asked to.

Boston Dynamics unveiled a new hand for Atlas on October 1. It has four fingers and 13 degrees of freedom, up from seven on the previous hand, and it is directly actuated. The company calls it the companion to the current Atlas: an agile humanoid it says can carry more than 100 pounds, built with the simplicity needed for mass manufacture. The film on this page is the product introduction. The design note is on the company’s blog the same day.

The old hand was for grasping a wide set of objects. This one is for manipulating them. A four-degree-of-freedom thumb opposes three fingers of three degrees each. Fingers can splay. The thumb can slide along and across the other digits. Boston Dynamics lists pinch and tripod grasps, in-hand reorientation, recovery from a slipping object, and triggered tool grasps: drills, torque drivers, grinders, nail guns, and welding torches. The pinkie is gone on purpose. A fifth finger, the company said, would add actuators, cost, volume, and failure points.

Actuation is the other half of the argument. Motors sit in the joints. One actuator type is used throughout. There are no cables crossing joints. Boston Dynamics says that choice keeps the hand close to simulation, which is what sim-to-real reinforcement learning needs, and keeps it repairable. Dense pressure sensing covers fingertips and palm. Backdrivable actuators supply proprioception. Strength, the company says, stays near the previous hand.

We treat the October 1 post as a real hardware announcement, not a render-only concept. We can say the earlier Atlas hand could do striking grasps and was never meant for production in the tens or hundreds of thousands. The new hand is the manufacturable answer Boston Dynamics is offering. It is still a company demonstration. No customer line, shift count, or unit price is in the launch note.

The film shows the hand on Atlas doing the grasps the blog describes. Four fingers, direct drive, no tendon run across a joint. That is a different architecture from the tendon hands circulating this month. Boston Dynamics is arguing for fewer degrees of freedom, one motor type, and tool use, not for a replica of a human palm. Independent coverage confirms the spec sheet. It does not confirm a fleet already using the hand on a factory shift.

Atlas now has a four-finger, 13-degree hand, announced October 1, directly actuated and aimed at tools and in-hand work. The 100-pound-plus body payload is Boston Dynamics’ figure for the robot, not a measured hand capacity. Bar four reflects a named product from a company with a public design note and outside coverage, not a completed deployment.

MindOn posted a Mind-1 demonstration at the end of September built around speed, not a new body. The film moves from packing and folding clothes to flipping parcels and pulling garments off a sofa. The same model is shown on dual-arm stations, a wheeled dual-arm, and a humanoid. On-screen labels read real time and fully autonomous. Those labels are the company’s. No outside lab has published a clock on the same runs.

The argument in the caption is that finishing a task is no longer the interesting line. A robot that folds a shirt in several minutes does not replace a person on a packing line. Mind-1 is presented as the step that keeps manipulation near human pace across more than one machine. Parcel handling, shirt folding, and room pickup are the examples. They are short, selected clips, not a shift log from a warehouse or a home.

Cross-body transfer is the other claim. If one policy can run on a fixed dual-arm, a wheeled dual-arm, and a humanoid, the model is doing more than a single rig’s demo. MindOn has not published the training mix, the success rate, or how often a human reset the scene between takes. Humanoids Daily, covering the September 30 film, noted the same limit: pace is asserted on the tape, not measured by a third party.

Laundry and parcels are a crowded demo shelf this month. Other firms have shown uncut washer-to-shelf runs on wheeled torsos. Mind-1’s distinction, as MindOn draws it, is one model at working speed on three morphologies, including a biped that retrieves clothes from furniture. Viewers can judge the pace on the film. They cannot judge uptime, wrinkle quality, or parcels per hour from the cut alone.

The clip belongs with the other two tiles only as a manipulation story. Origami was a scored paper task on a shared hand at IROS. Atlas is a new hand on one humanoid. Mind-1 is software shown on several robots, with speed as the headline. No factory customer, no price, and no endurance hour is attached. Bar zero matches a vendor demonstration.

Mind-1 is MindOn’s manipulation model, shown packing, folding, flipping parcels, and retrieving clothes, on dual-arm and humanoid hardware, at a pace the company calls near human and fully autonomous. The film is the evidence. A timed third-party trial is not.

Agility Robotics unveiled Digit 5 on September 15 as its first humanoid engineered for cooperatively safe work at scale. The company says the robot can work in close proximity to people without the physical safety barriers traditional automation still uses. The product film on this page is Agility’s official introduction. The firm frames Digit 5 as a warehouse and factory tool built on a decade of research and years of earlier Digit deployments, not as a show-floor mascot.

Cooperative safety, in Agility’s wording, means the robot watches for people and then avoids them, stops, or sits. Proprietary detection software and an independent safety controller are supposed to make those moves without a fence. Visual and audio cues tell nearby workers what the machine intends to do. IEEE Spectrum and Ars Technica both reported the sit-down behavior as the blunt answer to a falling biped: put the load down and become a stable seat before a person is close enough to be hit.

The published hardware numbers sit next to that claim. Agility lists a 23-kilogram, or 50-pound, payload and a reach of about 2.1 meters. A 90-minute battery with a nine-minute charge is how the company argues more than 20 hours of work in a day. Spectrum described a 1.8-meter, 129-kilogram machine trained, Agility says, on 65,000 hours of real-world operating data. Those figures are Agility’s. They are not a third-party shift log from a Digit 5 line that is already running without cages.

Spectrum also noted the hardware is real and that Digit 5 is not yet deployed in factories. That sentence matters more than the launch adjectives. Earlier Digits worked behind workcells. Digit 5 is the generation Agility says can leave the cell. Until a customer site publishes hours with people walking the same aisle and no fence, cooperative safety remains a designed behavior on a new body, not a finished industrial standard.

The film shows gait, reach, and the safety language Agility wants attached to the machine. A conference hall in Pittsburgh is not a Digit 5 factory floor. Agility’s own site is the source for the phrase “first humanoid engineered for cooperatively safe work at scale.” Independent coverage confirmed the launch. Factory-scale work beside people without barriers is still the claim being tested.

Digit 5 is Agility’s fifth-generation warehouse humanoid, announced in mid-September and built to detect people and then avoid, stop, or sit instead of living behind a cage. Payload, charge time, and the 65,000-hour data line are company figures. Work beside people without barriers is the promise still being proven in the field.

On Relentless, 1X founder and CEO Bernt Børnich sat with Ti Morse and said the company is preparing to manufacture and ship 50,000 NEO humanoids in 2027. That number is a target from the September 18 interview, not a completed order book. Børnich’s own caution came in the same opening stretch: the hard part is not only shipping 50,000 units, but making sure they do not come back.

The factory math he gave is also a plan. The Hayward, California line is built for about 10,000 units a year at full ramp. He said 1X will not hit 10,000 this year because the line reaches that rate late in 2026. A second site in San Carlos is being built for about 100,000 units a year, with most of that volume arriving late in 2027. Combined nameplate capacity is not the same as 50,000 robots walking out the door next calendar year.

Børnich treated production as harder than prototypes. Rare faults start to matter when output moves from hundreds of units into the thousands. He described quality gates, a roughly four-week path from a major CAD change to a walking robot, and a preference for small teams and fewer parts than a car. Those are manufacturing arguments from the founder. They are not third-party yield numbers from either plant.

The rest of the conversation is product philosophy, not a shipment report. Living with NEO, automating household tasks, warmth in the design, world models, teleoperation, and the line that 1X is “diversity-bound rather than data-bound” are how Børnich wants the home robot understood. Formal safety work before broad home use is the limit he put on the consumer story. Treat that as stated intent.

The interview was recorded at 1X and posted in mid-September. It is a California manufacturing conversation. It does not place NEO on the IROS exhibit floor in Pittsburgh. The 50,000-unit figure belongs to Børnich’s plan for 2027, not to a count of robots already shipped.

1X is targeting 50,000 NEO shipments next year, backed by a 10,000-unit Hayward ramp and a planned 100,000-unit hall in San Carlos, gated by quality and returns. NEO remains the home-robot platform the company has been building toward. Fifty thousand units is next year’s goal, not this year’s tally.

JAKA Robotics posted a recap of JAKAπ at the World Humanoid Robot Games: dancing, competing and taking on new challenges on a compact frame. The tape is a highlight reel, not a medal table. JAKA is a Shanghai cobot maker that spent more than a decade on collaborative arms, motion control and joint modules before putting that stack into a child-sized humanoid. π is that first body. The games clip is how the company wants the jump from factory cobots to an embodied platform to be seen by labs and show floors that already know the arm line.

JAKAπ stands about 122 centimeters and weighs about 42 kilograms, with 27 degrees of freedom and a listed 3-kilogram arm payload. JAKA sells it as an open development machine for classrooms, retail, entertainment and companion settings, with SDK access, voice control and a dual “fusion brain” on an Intel platform that splits high-level reasoning from real-time motion. Walk, run and jump claims sit in the spec sheet. The games footage is the public proof file for balance and choreography, including a Tai Chi showing that JAKA itself billed as a top-ten finish at the Games.

That cobot history is the part of the caption that is not just a montage. Joint design, force control and safe motion next to people are what JAKA already ships in plants. A humanoid that dances on a stage is using some of the same joints under a different costume. Whether those joints hold up in a classroom or a shop floor for a full shift is a different test than a games highlight cut to music. The company is asking viewers to read the decade behind the reel, not only the performance on top of it.

The World Humanoid Robot Games have become the place Chinese vendors take a new biped to be seen walking under lights. JAKA’s own site framed π’s Beijing appearance as validation after a June 2026 public debut. A recap video weeks later is standard vendor aftercare: keep the robot in the feed after the arena lights go down. It is not an independent scorecard of how π placed across every event, and it does not replace a timed walk or a classroom log.

Watch the clip for what it is. Coordinated whole-body motion on a short humanoid. A cobot company showing it can put a biped on a floor and make it perform in front of cameras. No payload trial, no classroom study, no endurance hours. The open SDK and the listed markets are how JAKA says the same machine leaves the arena and enters a lab or a store. Those paths are still product language until someone publishes a semester or a shift with the unit still standing.

Write π as a compact JAKA humanoid with a games reel and a cobot pedigree. Height, mass and DoF are the company’s numbers. The Tai Chi top-ten line is JAKA’s own result posting. The tape D5z-MqDOuGo is the look-back they circulated after the Games. Until a third party times a walk or logs a classroom week, this stays a vendor recap. Do not read it as a LinkerHand wrist or as an IROS booth.

X-Pro is a wheeled quadruped that the vendor dated to an August 2026 model launch and is now circulating as a patrol and inspection platform. The pitch list is security patrol, fire rescue, facility inspection and oil-and-gas rounds. Wheels on the legs are the product idea: roll on flat ground, walk when the path breaks into stairs or gravel. The tape is a demonstration of that hybrid gait and of cameras doing a route, not a signed contract with a refinery, a plant or a firehouse.

Wheel-legged dogs have become their own shelf this year. Faraday Future’s Aegis family sits on the same idea at a different price. X-Pro is being sold into the industrial inspection slot: perimeter loops, plants, and sites where pavement and broken ground show up on the same shift. Vendor pages tied to the August clip describe autonomous patrol, navigation, obstacle avoidance and recognition, with a note that the footage is for demonstration. That disclaimer belongs in the same paragraph as the mission list, not after the sale.

Fire-rescue and oil-gas language is the high-risk end of the brochure. A dog that can carry thermal, gas or camera payloads is a familiar quadruped story. Putting wheels on it is how the seller argues speed on pavement without giving up a curb or a stair. None of that is a certified firefighting appliance or a documented leak walk at a live well pad. It is a configuration looking for a buyer who already runs patrol robots and wants mixed terrain in one chassis instead of two.

The August date is why the clip is still moving in late September. A summer launch reel often gets a second pass weeks later while other stories fill the same brief. X-Pro is not a biped and it is not a classroom kit. It is a wheeled dog with a job list. Keep the three tiles in separate boxes. L30 is a tendon hand. π is a Games recap from JAKA. This is a patrol platform on four legs and four wheels.

What can be stated from the listing and the video is limited and clean. A new X-Pro model was presented in August 2026 as a wheel-legged robot dog. Intended work: security, inspection, fire-rescue support and oil-gas patrol. The file TnpouUKpMFQ is the demo attached to that launch line. Capabilities shown are motion and sensing on a course the seller chose. Treat payload kits, runtime and weather limits as unspecified until a datasheet, not the title card, puts numbers on them.

Do not upgrade a product video into a deployment. No city, plant or brigade is named in the copy we were given. No price. No fleet size. The honest brief is a hybrid quadruped offered for patrol-shaped work, dated August, still being pushed in September. If a site logs nights with the unit on a fence line, that is a later tile. This one is the launch tape and the job list on the tin.