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.

The team behind Asimov 1 has open-sourced the robot’s locomotion policy and the training code that produced it. The drop includes a trained policy checkpoint plus Isaac Lab training code: reward definitions, hardware-tuned actuator configurations and domain-randomization settings. The clip starts at ten seconds, where that stack is laid out. The point is a starting walk other labs can fork, not a finished general controller for every pair of legs now leaving a factory in China or California. Asimov 1 itself is the open-hardware biped this policy was trained to move.

A locomotion policy is the controller that turns balance, step timing and contact into a walk. Publishing the checkpoint means someone else can load the weights instead of training from noise. Publishing the Isaac Lab recipe means they can see how those weights were grown: which rewards were counted, how the actuators were modeled on the real motors, and how the simulator was randomized so the policy would not overfit a single floor or a single mass. Without those three files together, a public “walk” is usually just a highlight reel.

That last piece is the useful one for anyone who does not own an Asimov 1. Domain randomization and actuator settings are where a simulated walk usually dies on hardware. By putting those files next to the checkpoint, the release is asking other groups to retune the same loop for a different gearbox, a different mass, or a different gait, rather than rebuild the scaffolding before they can take a first step in their own lab. That is the difference between a demo video and a repo someone can actually break.

The note that came with the video is modest on purpose. Adapt the training setup when the hardware changes. Try another walking style. Build toward a new skill on top of a walk that already exists. It does not claim the policy transfers zero-shot onto a Unitree, a Figure or Faraday Future’s classroom Mini. It claims a foundation that a graduate student can clone without first inventing the reward book. Read it as an invitation to train, not as a guarantee that the next robot will walk on Friday.

Open locomotion stacks have become a second beat beside product launches this year. A $9,990 humanoid is a price on a purchase order. A public Isaac Lab walk is a tool in a repo. One asks a school district to write a check. The other asks a researcher to break the gait on Monday and send a pull request by Friday. Both belong on the same brief. They are not the same kind of news. One is a catalog. One is a method other people can reuse.

The facts in the tape and the caption are the release itself: Asimov 1 policy weights, Isaac Lab training code, rewards, actuator configs, domain randomization. No third-party benchmark or cross-robot transfer study is attached. Treat it as an open walk from the people who built the reference hardware, and then watch who actually trains the next policy on top of it. Until that second paper or fork shows up, the story stops at the files they posted.

Qiyuan Robotics, sold in English as PrimeBOT, held a 20 September launch for two personal machines: the Qiyuan Q1 and the Qiyuan T1. Both standard trims open at 19,999 yuan, about $2,800. The tape starts at twenty-eight seconds. These are home-size robots, not factory bipeds. The company is selling companionship, a custom shell and a follow-cam, with Tencent Cloud WorkBuddy listed as the shared layer for talk and tracking across both bodies. Shanghai was the main stage; other cities ran experience events the same week.

Q1 is the small humanoid. Height is about 88 centimeters. Weight is about 15 kilograms. The body folds into a backpack or carry-on, and the shell can be customized, including 3D-printed parts in the launch pitch. The brief positions it as family company and a street-level gadget, not a research platform. Explorer Edition is listed at 26,999 yuan. That is the trim for buyers who want the same foldable frame with a thicker software and accessory pack. Voice, personality and motion are part of the customization story Qiyuan is selling with the hardware.

T1 is the transformable unit. Standing height is about 100 centimeters. Weight is about 16 kilograms. It switches between wheeled-humanoid and quadruped modes, and it is sold with a camera-follow and camera-move job for indoor-to-outdoor shooting. Pro is 29,999 yuan. Where Q1 is a packable figure you can put away, T1 is the one that changes stance and rolls when a walk is the wrong tool for the hallway or the park path. The company is pitching that change of form as the reason to buy T1 instead of a second Q1.

What the two share is the cloud layer. Both are integrated with Tencent Cloud WorkBuddy for multimodal interaction and intelligent following. That is the home-use argument: a voice, a trail, a small body that does not need a lab bench. It is also a vendor stack. WorkBuddy is Tencent’s. The robots are Qiyuan’s. The launch does not publish a detailed offline-fallback spec for when the network drops in an apartment. If the follow feature needs the cloud, that limit belongs next to the price, not in a footnote after a return.

Orders are set to ship in sequence from 1 October, first come first served. That is a calendar claim, not a tracked first-wave of boxes on doorsteps. A starting price of 19,999 yuan puts both machines in the same consumer band as other compact home humanoids now printing list prices in China, while Faraday Future’s Mini is the dollar figure sitting on the lead tile of this brief. Pre-order channels cited around the launch include the official mini-program and major domestic ecommerce storefronts.

Write the pair as listed, not as a household census. Q1: 88 cm, 15 kg, folds, 19,999 / 26,999 yuan. T1: 100 cm, 16 kg, wheeled-to-quad, 19,999 / 29,999 yuan. Shared WorkBuddy. Ship date 1 October, according to the company. The clip is the launch tape from the Shanghai event, not a week of homes already living with either machine. Until those October boxes land and stay powered, this remains a priced announcement with two distinct bodies on the same receipt.

A hexapod teaching kit now ships locomotion, a small arm and a spoken-command stack on one chassis. The listed hardware is an 18-degree-of-freedom body running inverse kinematics; omnidirectional crawl, climb and dance; a two-degree-of-freedom arm for grab and carry; a two-megapixel camera for face, gesture, pet and line tracking; WonderLLM and ESP-Claw for speech; a six-axis IMU; and programming in Arduino, Scratch and MicroPython, with PC software and a phone app for direct drive.

Eighteen degrees of freedom on six legs is the standard three-servo limb. Inverse kinematics keep the body level while the gait changes, which is what lets a hexapod look like it is walking instead of ticking. The two-axis arm is the extra that turns a crawler into a machine that can pick something up and move it. Camera jobs match a first vision lab: find a face, read a gesture, follow a painted line, notice a pet in the frame.

WonderLLM is the branded module for talking to the robot. ESP-Claw is the on-device agent layer, so sensing, a decision and an action can run on the board instead of waiting on a cloud round trip. That pairing is the 2026 hook on an otherwise familiar hexapod recipe. The IMU exists so the chassis knows when a leg has left the ground, or when a student has picked the robot up mid-stride and put it down facing the wrong way.

The software path is built for schools and clubs rather than a factory cell. Scratch and the phone app cover beginners who want motion without writing a gait from scratch. Arduino and MicroPython cover students who want to own the claw, the camera loop and the walk. The pitch is one machine that can crawl on Monday and take a voice command on Friday without a second piece of hardware.

The video is a product demonstration of those features in sequence. It does not include a classroom trial with scored lessons, a published battery figure, or an endurance run that would tell a teacher how long the servos last in a noisy lab. Treat it as the spec list made visible, starting at the fifty-three-second mark where the hardware card is read aloud.

What you have, then, is a six-leg teaching platform with a small manipulator and an on-chip chat stack. It is not a full-size humanoid and it is not sold as one. It is a way for a student to hear a robot answer and then watch it close a claw on a block, which is how a lot of people will first meet embodied software this year.

Content creator Frankie LaPenna fought a six-foot humanoid styled as a T-800 at a private cage match organized by REK Robotics. The video is that bout. LaPenna, in heavy pads, lands punches on the metal. The machine answers with kicks that drive him into the fence and then onto the mat. The announcer frames it as a person stepping in against a six-foot robot. The clip is short, loud and built to travel.

REK billed the night as the first human-versus-Terminator fight and a turning point. That wording is the promoter’s. Similar man-versus-machine exhibitions have been staged before, which is why the “first” belongs in quotation marks. The 18 September event in San Francisco was organized to promote a virtual-reality game in which people pilot robots against one another or against a live opponent in a cage.

The robot is teleoperated from backstage, not choosing its own kicks. Coverage of the bout has identified the body as a modified EngineAI T800 under the Terminator shell and glowing eyes. Kick-force numbers that appeared in entertainment write-ups have not been published as a laboratory measurement. What the footage shows is a remote pilot, a heavy chassis and enough impact to move a padded adult across the floor.

LaPenna is a stunt and prank creator rather than a ranked fighter. After the match he showed a swollen hand from hitting metal, which is the injury detail that holds. REK’s chief executive later told an outlet that LaPenna came back and finished the robot once the viral kick sequence was over. The clip most widely shared still ends with him on the floor. Both accounts can sit in the same week without turning the night into a sanctioned result.

Pads, a cage and a pilot in a headset make this a show with real contact, not a title fight and not a new combat humanoid offered as a product. The thing being sold is the teleoperation loop and the game around it. REK has already said more bouts are coming and that the machine in this clip was not the last one in the lineup.

The facts that survive the edit are simple. The date is 18 September. The promoter is REK. The opponent is a six-foot T-800-styled humanoid driven by a person. The man in the cage is Frankie LaPenna. The tape is the exchange of punches and kicks from that private match, not a factory demonstration and not an autonomous fighter working on its own.

Unitree put Dex5-S on the channel on 21 September: a precision biomimetic hand, human-scale, from $6,500 before tax and shipping. All 22 joints backdrivable. Each joint listed with limit impact-torque protection. Domestic coverage puts the floor at 39,900 yuan. The 43-second clip is the announcement tape. It is a product drop, not a Games medal and not a factory cell. The number that moves the market is the sticker next to a 22-DoF count most labs still pay five figures for.

Launch write-ups describe five fingers at roughly 1:1 hand size, about 620 grams in one recap, thumb and pinky at five degrees of freedom and the middle three at four. Dual encoders and backdrivable motors are the compliance pitch — an external shove can move the joint instead of fighting a locked gearbox. That matters for reinforcement learning and for not snapping a wine glass. Unitree also flags a Pro SKU with tactile sensing versus a Standard without. Spec sheets are still thin on fingertip force and cycle life.

Dex5-S sits above Dex5-1 in the family tree. The prior hand was described as 20 degrees of freedom with a 16-active / 4-passive split and a large tactile array. The new unit’s claim is every joint driven and backdrivable. Do not flatten that into “more sensors.” It is more actuated axes plus a price that undercuts Allegro, Shadow and most direct-drive research hands. Whether it undercuts OmniHand or Wuji on a timed grasp is untested here.

The video drew the usual respect for showing metal instead of only renders, and the usual “that is not how you open a can.” Fair on both. A $6.5K 22-DoF hand will land on G1 / H1 / H2 arms because that is Unitree’s stack. Integration hours will decide if the DoF count is usable or just countable.

Bar at three. A named SKU, a public price, and a backdrivability claim across all joints is more than a concept reel. It is not a third-party durability log. Three bars means the product exists and the test has not.

Watch for a force table, a Pro versus Standard split you can order, and a humanoid wearing two of them through a shift. Until then Dex5-S is the cheapest ticket into the 20-plus-DoF club Unitree has sold.

XPENG’s new IRON clip is a real-world interaction test, labeled as such. Four capabilities: autonomous orientation toward the speaker, dynamic identity memory, seamless multilingual switching, multi-intelligence Q&A. The caption says IRON recognizes you, understands you, responds to you, powered by Physical AI. The footage is three testers playing car buyers in the robotics lab. That framing matters. This is not a dealership log. It is an R&D demonstration of the social stack sitting on the production-line body XPENG walked off the line earlier this month.

In the identity beat, one tester borrows another’s nickname; IRON corrects him and recalls vehicle interest. The company says recognition holds when the three swap places and clothes. Memory is described as face plus voiceprint plus conversation history, updated on later encounters. Language switch is presented without a separate “now speak English” command — Mandarin and English in the exchanges. XPENG credits a multilingual foundation model and balanced training data. No published accuracy, latency or overlapping-speaker score accompanies the reel.

Hardware context from the September production announcement, not this video: 76 body degrees of freedom, 21 per hand on the production revision, three Turing chips and a company rating of up to 2,250 TOPS onboard. Mass production targeted by end of 2026, commercial deliveries talked for 2027. Those dates stay on the factory slide. The interaction tape does not prove them.

Physical AI is XPENG’s corporate banner — cars, IRON, flying boxes on the same Turing story. VLM conversation and VLA action are the split the company has used since AI Day. This clip is almost all VLM: look at the person, remember the person, answer the person. Walking and grasping are off-screen. Do not write that IRON now sells cars.

Useful as a humanoid-as-clerk sketch. Thin as evidence. Three scripted buyers in a lab is how you start a memory demo. It is not how you measure a showroom.

Bar stays at zero. A labeled R&D interaction test belongs on the page beside Helix’s scored homes and Unitree’s priced hand. It does not get a field rating until XPENG publishes a recognition table or parks IRON on a real floor with real shoppers.

Dobot’s X-Trainer explainer is back in circulation: an advanced dual-arm teleoperation and AI training platform built around two Nova 2 cobots. The pitch is higher education, research and technical training — capture repeatable demonstration data without welding a rig from scratch. Imitation learning, computer vision, human-robot interaction. The chapters walk introduction, teleop, why the data matters, lab pain, hardware, SDK, curriculum. It is a product video for a kit that has been on the market since 2024. It belongs here because physical-AI labs keep buying dual-arm data, not because it walks.

Hardware is two 6-axis Nova 2 slaves plus two 6-DoF master hands with end handles. Nova 2 repeatability is listed at ±0.05 millimetres. Combined dual-arm span is about 1,200 millimetres; single-arm reach 625 millimetres. Collision detection is five-level; the cobots carry ISO 15066 collaborative certification. Dual e-stops sit on the table. Master-hand load on the operator is advertised under 400 grams. That is the safety case for putting undergraduates on the handles.

Software is DobotStudio Pro plus an API that exposes master joint angles, lock/sync, slave Cartesian control, gripper stroke, and RGB/depth frames. Record starts from a green button on the handle. GitHub’s embodied-dobot/x-trainer stack maps the leader into Isaac Lab, VR (Quest/PICO) and a 30-hertz HDF5 pipeline aimed at VLA training. Dobot has claimed a 70 percent cut in scenario training time, down to about two hours for simple tasks. That figure is marketing from the 2024 launch. Leave it labeled.

Why a humanoid page cares: most foundation-model teams still collect bimanual data on tables, not on bipeds. X-Trainer is the industrial-grade version of that table. Students get collision-limited cobots instead of a research arm that can break a wrist. Researchers get cameras and a leader that already talks to LeRobot-style formats. It does not replace a Digit or a TRON 2. It feeds them.

The clip is not news of a new robot. It is news of the channel still selling data infrastructure while the rest of the week argues about hands and policies. Fair placement as digest two. Do not write “humanoid trainer.” Write “dual-arm cobot classroom.”

Rating stays at zero. A shipping education kit with a new explainer is useful context, not a field rating. Watch for a published student dataset or a paper that names X-Trainer as the capture rig. Until then it is two Nova 2s and a pair of handles doing what they were built to do.

Linkerbot’s 19 September short is a speed argument, not a launch. LinkerHand L30 is tendon-driven, high degree of freedom, and fast because the actuators sit behind the fingers. Little mass left to accelerate. That, the caption says, is what coordinated work at this joint count requires. Footage from earlier this year is cut with more recent takes. The title on the channel calls it the fastest high-DoF hand they know of. That superlative stays in the video. It does not become a measured ranking here.

The product manual is the sheet we will use. L30 is a modular anthropomorphic hand from Linxin Qiaoshou / Linkerbot. Transmission is UHMWPE tendons in PTFE tubes, motors remote from the digits. Coordinated control of five fingers: 17 degrees of freedom, 21 joints (17 active, 4 passive), including a wrist degree of freedom on that count. Control interface CAN FD. Reseller pages scatter 21- and 22-DoF labels; treat those as marketing variants of the same architecture, not a second product.

Published performance numbers that repeat across decent listings: about 1.4 kilograms, ±0.20 millimetre repeatability, fingertip forces in the low-newton range on the flagship tendon version, core motion advertised near 400–450 degrees per second. Payload figures disagree between shops — 2 kilograms on one sheet, much higher on another — so this brief will not pick a load. Speed and remote actuation are the claims the new clip actually illustrates.

Why tendons again after a week of Wuji direct-drive: different packaging. Direct drive puts a motor at the joint and accepts heat and distal mass. Tendons move the mass into the palm or forearm and accept stretch, routing and maintenance. L30 is the second bet. If the cut-together reel is honest, the fingers look light because they are. If the “fastest we know” line is just editing, it will show the first time someone times an open-close against Wuji Hand 2 or OmniHand on the same task.

Linkerbot already had an L30 on the market before this montage. The news is the speed framing plus an admission that some shots are old. That honesty helps. A lab buying a high-DoF tendon hand still has to ask about tendon life, calibration drift and whether the wrist DoF is in the SKU they will receive.

Rating stays at zero. A tendon flagship with a new cut is the hardware counterpoint to TRON 2’s Wuji integration and Dobot’s classroom arms. Three ways to get fingers on a task this week. None of them is a counted line. Watch for an L30 on a named humanoid, timed, not montaged.

Apex Hand, Dexcel Robotics’ high-DoF five-finger unit, is on a clip that is more vendor tape than medal table. The company says it was the only high-DoF hand used across both high-load and precision-manipulation challenges at the World Humanoid Robot Games 2026. That is the caption, not an organizing-committee roster. Independent coverage of the new eight-event hand track still runs through AGIBOT’s OmniHand. Treat Apex’s line as a promotional case shot against a Beijing backdrop.

The sheet Dexcel and Rysen publish is consistent across product pages. Apex lists 21 degrees of freedom with 16 active, about 1.2 kilograms, fingertip force above 20 newtons, and a 30-kilogram maximum-load claim. Joints are sold as backdrivable. Skin is sold as full-hand e-skin. Repeatability is under 0.5 millimetres. Founder copy describes a tendon-driven hybrid, not a pure direct-drive finger. Those are spec-sheet facts. They are not a start list from the Oval.

What the new clip adds is the two-lane pitch. One body is supposed to take load and still do fine work — tools and tweezers without a claw swap. The Games, 22–26 August at the National Speed Skating Oval, were the first edition with a dedicated dexterous-hand track. That is the story Dexcel wants: dexterity, strength and reliability tested next to competitors. It is also the story that is cheapest to cut with stadium b-roll and hardest to verify without a named heat.

Dexcel has spent a year moving Apex from launch reels toward public rooms. Smartphone use and a Kapandji score were the early beats. This tape is the next one: stop talking about a hexagon of specs and start talking about the Games. Fair enough as a marketing sequence. It does not retire the question every hand still owes: how many cycles, how many resets, how many fingers swapped per shift.

For industrial readers the comparison is OmniHand, not a parallel-jaw that already packs boxes. OmniHand’s argument is a published Games result. Apex’s argument is one end effector covering heavy and fine. Both can be true and still be early. A hand that holds 30 kilograms on a sheet is a useful scene. A hand that held a scored run in Beijing is a different product. Dexcel has not published the second number.

Rating stays at zero. A high-DoF hand claiming both load and precision at the Games is worth logging next to policy tapes like OM-1. One is trying to look transferable. The other is trying to look unbreakable. Neither is a deployment metric. Watch for a named team and a named event, not a second take of the same fingers on a clean floor.

WUJI is on a clip that is more origin story than product launch. The company says it takes only a few seconds to see a dexterous hand come to life, and that those seconds sit on seven years — motor to direct-drive hand, small factory to ICRA. That is Wuji Tech’s line, not a third-party audit. The official file runs twenty-three minutes under From Motor to Dexterous Hand. Treat the short hook as the sell. Treat the runtime as the history they chose to publish.

The hardware WUJI published is consistent across the site and docs. Wuji Hand lists 20 active degrees of freedom, four per finger, serial direct-drive joints, a skeleton of about 580 grams, 15 newtons at the fingertip and 10 kilograms static grasp. Hand 2, shown at ICRA 2026 booth 121, adds back-drivable actuation, force-position control and a 1-kilohertz loop. Those are spec-sheet facts. They are not proof the hand held a shift after the booth closed.

What the new clip adds is the no-playbook pitch. From data collection to end execution, Wuji Glove and Wuji Hand are supposed to be one stack: capture, stream, act. Direct drive is why. A tendon map hides slack. A joint the controller owns is what a learning pipeline can write to. That is the story WUJI wants. It is also the story that is easiest to narrate over a wake-up shot and hardest to keep when heat and dust show up.

WUJI has spent 2026 moving from factory trials toward public rooms. ICRA was the work-shaped sibling of the first Hand launch — thermal headroom and models in MuJoCo, not only a pretty curl. This week’s film is the next beat: stop talking about a motor and start talking about seven years. Fair enough as a marketing sequence. It does not retire the cycle-count question every hand still owes.

For a humanoid page the comparison is Apex and OmniHand, not a gripper that already ships. Digit’s argument is a cell coming down. WUJI’s argument is a light 20-DoF palm a policy can instrument. Both can be true and still be early. A booth that makes fingers twitch is a useful scene. A line that keeps those fingers on for a week is a different product. WUJI has not published the second number.

Rating stays at zero. A seven-year direct-drive path is worth logging next to OM-1 and Apex. One is trying to look employed as intelligence. The others are trying to look honest as hardware. None is a deployment metric. Watch the clip, keep the 20-DoF sheet, and wait for a buyer with hours. If the next video is another wake-up, this stays a company hardware story — which is still news. It is just not a counted fleet.

ALICE M1, AeiROBOT’s wheeled mobile humanoid, is on a clip that is more kinematics than marketing. The robot carries an object to a target pose by driving position and orientation together while both hands stay on it. The two arms and the load are treated as one closed kinematic chain, so the relative pose between the hands does not drift. That is the point of the tape: keep the grasp from shearing the part while the whole upper body walks the object through space.

Once the part sits at the target, ALICE M1 self-motions. Self-motion here means using redundant degrees of freedom to change shoulders, elbows, waist and knees without moving the object or the hands. The load stays put. The skeleton rearranges around it. In a tight aisle that is how you unload a joint that is near a limit, or slide a shoulder off a shelf edge, without putting the box down and picking it up again. The clip is a lab demonstration of that idea, not a warehouse scorecard.

The platform under the arms is the M1 already on AeiROBOT’s sheet: height 130 to 180 centimetres, about 97 kilograms with the base, 31 degrees of freedom, omnidirectional wheeled drive rather than a biped gait. Arms are 7-DoF; hands 6-DoF. Compute options include NVIDIA AGX Orin. That wheeled base is why a closed-chain carry is a fair demo. The robot is not also fighting balance on two feet while it solves the chain.

Closed-chain bimanual control is old theory and still rare on a product page. Most humanoid clips show two arms that happen to touch the same box. This one states the constraint: relative hand pose held constant, then null-space motion for posture. If the implementation is as clean as the caption, it is useful in aisles and cages. If the object still wriggles off-camera, it is a slide. The video is the evidence. There is no independent timing study attached.

For industrial readers the useful sentence is the last one in AeiROBOT’s own copy: keep the grasp and keep working when the room is cramped. That is a real failure mode on mobile manipulators — the part is fine, the elbow is not. Self-motion is how you buy another centimetre without a regrasps. Whether M1 does that at speed, under mass, next to a person, is not in this reel.

Rating stays at zero. A kinematics demo on a wheeled humanoid is worth logging next to service-floor tapes like A3 Ultra. One is trying to look employed. The other is trying to keep two hands honest. Neither is a deployment metric. Watch for a timed carry through a real fixture, not a second take of the same box on a clean floor.

VOGEBOT is a Chinese quadruped that puts four wheels on four legs and a person on top. The clip making the rounds this week — including a Portuguese-language tech recap — treats it as a mobility experiment, not a humanoid. Wheels for the smooth bits, legs for the ruts, a saddle instead of a gripper. That is a different product family from A3 Ultra and ALICE M1, and it belongs here only as the rideable-robot beat that keeps colliding with humanoid news feeds.

Public write-ups describe a wheeled quadruped that can take a rider across slopes and uneven ground. Independent spec sheets are thin. Do not borrow numbers from DaxAI’s Qiji horse or Unitree’s B2-W just because those also carry a person. VOGEBOT is its own tape. What we can say from the video and the caption is the layout: four legs, four wheels, one rider, rough terrain as the selling point. Speed, mass, range and maker legal name are not in a document we will treat as locked.

Rideable quadrupeds are having a second life in China after Unitree showed a person on B2-W and DaxAI showed a full-size robot horse. VOGEBOT sits in that lane: personal transport that can step where a scooter cannot. Inspection and rescue get named in recaps because every rideable robot eventually does. Those uses are speculation until a named agency buys one. For now the verified object is a demo machine that carries a human and mixes wheels with legs.

Why it is in a humanoid brief: readers keep asking whether the future of “robot mobility” is two legs at human height or four legs under a seat. Digit and A3 Ultra bet on walking through a door. VOGEBOT bets on not needing the door. Both answers can ship. Only one of them folds towels. The digest is here so the page does not pretend every interesting robot this week was a biped.

Safety is the open file. A wheeled-leg platform with a rider is a vehicle. Balance, braking, and what happens when a leg tucks at speed are not addressed in a highlight reel. Until there is a certification story or a test report, treat “carry a person” as a demonstration, not an invitation. Same standard we use for robot horses.

Rating stays at zero. A rideable hybrid is news; it is not a humanoid deployment. Watch the clip, keep the name distinct from Qiji and B2-W, and wait for a spec page with a company letterhead. If VOGEBOT publishes mass, speed and a buyer, it earns a longer slot. Today it is a mobility sidebar with a clean video ID and no invented numbers.

A hand with places to be: ETH Zurich’s Soft Robotics Lab put an anthropomorphic hand on the floor and taught the same fingers to crawl, steer, get up, type and shove. The paper, posted 15 September as arXiv:2609.17172, is “Fingers as Legs: Learning Self-Supported Locomotion and Manipulation with an Anthropomorphic Hand,” by Amirhossein Kazemipour, Hehui Zheng and Robert Katzschmann. The video matches it — untethered indoor and outdoor crawling, fall recovery, a keyboard, an object push — without a second locomotion mechanism bolted onto the palm.

The point of the work is reuse. A walking hand has to move its body, hold its own weight, and still interact with the world using the same digits. The lab kept the finger design and the position controller. Power and policy inference sit onboard, so the platform is self-contained rather than hanging off a tether. Reinforcement learning in a simulator calibrated from hardware measurements accounts for unequal fingers. In simulation their reward beat tuned quadruped rewards for speed. On hardware the skills are separate, task-specific policies, not one net that does everything.

Keyboard work runs without visual feedback. Object pushing uses an overhead camera. That split matters. It says the hand can stab keys while supporting itself from proprioception and contact, then accept vision only when the task needs a target in the room. Fall recovery is in the same reel as the crawl. A lab that shows the get-up is taking the joke form factor seriously instead of cutting away before the hand tips over on a rug.

This is not a warehouse humanoid and should not be written as one. It is a compact mobile manipulator that refuses to grow legs. Katzschmann’s group has spent years on tendon hands and soft-rigid hybrids; this paper is the locomotion half of that bet. mimic Robotics, the ETH spin-out, is the industrial cousin. The brief’s interest is narrower: fingers that walk still have to work when the floor is not a lab bench and the key is not centered under the fingertip.

What you can take to a humanoid page is the constraint. Most dexterous hands assume an arm is holding them up. This one does not. If a detached end-effector can crawl to a panel and press a key, the design space for “hand” changes. If it only works on the lab floor with task-specific policies swapped in for each trick, it stays a paper and a good clip. Either outcome is useful. Only one of them ships.

Rating stays at zero. An arXiv drop and a demo reel are research. Watch for a single policy that locomotes and manipulates without a swap, or a hand that does this under a real humanoid wrist. Until then: same fingers, own weight, crawl and type. That is enough for the digest. It is not a product, and it should not be filed next to Digit as if both were ready for a shift.

ROBROS released a clip of Real-Time Full Body Teleoperation on IGRIS-C, the compact Korean humanoid the firm has been showing since CES. Squats, kicks, turns, bows, box handling, bag punching, ball kicking — the operator’s motion maps onto the whole body in the moment, not only onto a pair of arms. The company line is that industrial tasks rarely live in the upper body. Locomotion, posture and manipulation have to run as one behavior. Real-Time FBT is their name for giving the operator that whole chain at once while the robot keeps posture and balance.

IGRIS-C is a 154-centimetre, 56-kilogram platform with 31 body degrees of freedom and tendon-driven five-finger hands. ROBROS has already posted dynamic whole-body motion — sit-to-stand, kicks, even a backflip — as onboard control. This week’s tape is the teleop layer on top: a person driving the same coordinated stack live. That is a different claim from “the robot can flip.” It is “a human can steer the flip’s cousins in real time.” The brief should not conflate the two. One is a policy. The other is a mapping.

Upper-body-only teleop is how a lot of humanoid data still gets collected. Legs walk on a policy; arms copy a person. Full-body mapping is harder because the operator’s balance is not the robot’s, and a squat that looks fine on a person can put a 56-kilogram machine on its back. ROBROS says coordinated posture holds through the clip. The clip is a demo floor, not a line. Treat it that way. A kick compilation is not a shift.

For industrial humanoids the useful sentence is the one ROBROS actually wrote: tasks need locomotion and hands together. A box lift that starts with a step and a squat is not an arm skill. If Real-Time FBT lets an operator record those sequences as one take, the data gets cheaper and the policy that learns from it has a chance of looking like work. If every kick still needs a spotter and a reset, it is a show with a new acronym.

Korea has been crowding the compact-humanoid lane. IGRIS-C’s pitch has been developer-friendly hardware plus imitation learning from teleop. Extending that teleop from chest-up to full body is consistent with the pitch. It does not, by itself, put a unit on a shop floor or answer Digit’s question about standing next to a person. One stack is for collecting motion. The other is for sharing a floor. Both can be true. They are not the same product.

Rating stays at zero. A real-time full-body mapping demo is worth watching next to Digit’s safety stack: one company is trying to stand next to people, another is trying to let a person stand inside the robot. Neither is a deployment. Watch for FBT used to collect a shift of work, not a kick compilation. Until that tape exists, IGRIS-C stays a capable compact humanoid with a better teleop story than last month.

UBTECH commissioned what it calls an industrial humanoid “super smart factory” in Liuzhou, Guangxi, over the 12–13 September weekend and spent the next days repeating two numbers: one robot off the line every ten minutes, more than 10,000 units of planned annual capacity. The official clip adds a third — a leap from 1,000 to 10,000 units in under nine months. Treat that last figure as company narrative about the capacity ramp, not a verified shipment log. What opened is a 14,000-square-metre plant, 13.8 metres high, built to make Walker S bipeds and Cruzr-series industrial machines.

Siemens Digital Industries Software is on the badge next to UBTECH. Digital twins and a manufacturing-operations layer are how the companies say they schedule a product with more than 2,000 screws of mixed spec. Cruzr Y1 and S2 units are described on depalletizing, palletizing, loading and haul. Collaborative arms, rotating tables and AGVs cover final assembly. A compact automated store of about 65 square metres is said to hold 112 finished robots. “Robots building robots” is the line Global Times and the company both used. Some of those robots on the floor are wheeled industrial platforms, not only bipeds.

A ten-minute takt at two shifts, year-round, is the arithmetic behind a 10,000-unit year. That is design cadence. It is not yield, uptime, or robots that leave the yard and stay sold. Independent write-ups this week have been careful on that point. UBTECH has already been selling industrial humanoids in smaller batches. Liuzhou is the attempt to make the next order of magnitude a factory problem instead of a lab problem. A commissioned plant is a real event. A quarter of continuous output at the advertised beat would be a different one.

The geographic choice is not an accident. Liuzhou is an auto and industrial town. A humanoid line that borrows car-plant software and logistics is the Chinese answer to the question Agility is answering with safety architecture: how do you stop building one-off machines. Capacity is the Chinese lever. Cooperative safety is the Oregon one. Both showed up in the same news week, which is why they sit on the same brief.

What the video does not show is a customer taking 10,000 Walkers. It shows a commissioned hall and a claimed beat. Coverage said the stock ticked up when the plant news hit. That is a market reacting to a factory ribbon. The industrial test is whether the line holds the ten-minute story after the cameras leave, and whether Walker S units from Liuzhou show up in the same factories that already run UBTECH pilots, with serial numbers attached.

Keep the rating at zero. A plant that can, in theory, print a humanoid every ten minutes is worth the digest slot. A plant that has done so for a quarter, at yield, with robots earning their keep on a customer floor, is the story that raises the bar. Count the machines that leave and stay sold. Do not count the machines the brochure says the clock can make.

Reward AI came out of stealth this week with OM-1, a general-purpose robot policy the company says was learned from human hands and then dropped onto other bodies. The slogan on the 14 September blog is one model, one data interface, any body. The video is a string of contact-rich bits — bartending, a nut coming off, Ethernet cables, phone packaging, laundry folding, tube sorting, a humanoid pick-and-toss. Technical notes live at rewardai.com/blog/OM-1/. This is a model launch, not a factory or a safety case.

The stack is called Omnibody. Capture is a wearable, the Omnibody Hand, a 7-DoF glove with tactile sensing, proximity, in-hand cameras and tracking. The lineage is Stanford DexCap work associated with Chen Wang, Karen Liu and Li Fei-Fei. Reward’s claim is that teleoperation rigs and parallel-jaw grippers dirty the data, so the right move is to record people working at human speed and train one policy. OM-1 is that policy. The same weights, the company says, run on tabletop arms, industrial manipulators and humanoids without a per-body fine-tune.

The sample-efficiency number they want quoted is short: a new long-horizon task from less than 30 minutes of human data. Zero-shot transfer across embodiments is the other headline. Neither has been independently bench-marked in public. The reel is selected. That does not make it fake. It makes it a vendor tape, which is how every foundation-model shop arrives. Watch the nut and the cable, not the cocktail pour.

Why it sits next to Digit 5 and a Liuzhou plant: manipulation is still the bottleneck after walking. Agility is adding skills on a body it already deploys. UBTECH is printing bodies. Reward is arguing the intelligence layer should not care which body you bought. If that holds even halfway, a warehouse can change grippers and frames without throwing the policy out. If it does not, OM-1 joins the pile of lab transfers that die on a second lighting setup.

Keep the human-data story precise. The model is trained on wearable demonstrations, not on months of robot teleop — that is the company’s distinction. “Human-level manipulation” in the caption is aspiration. What the clip actually shows is a set of short, staged tasks that look like the current frontier of contact-rich demo culture. Laundry folding and a tossed pick will travel. Nuts and Ethernet cables are the frames that matter if you work in a plant.

Rating stays at zero. A stealth exit with a blog post and a two-minute reel is not a fielded stack. Watch for a third-party robot running OM-1 on a shift, or a published comparison against the usual VLA baselines on a held-out body. Until then the useful sentence is Reward’s own, trimmed of the slogan: one policy trained on human hands, shown on several machines, claiming minutes of new-task data. That is worth tracking. It is not yet work.

GMO AI & Robotics, the robotics arm of GMO Internet Group, unveiled a van in Tokyo on 8 September and called it a humanoid ambulance. It is not for people. When a deployed biped goes down, the vehicle is supposed to roll out with an engineer, tools, spare parts and a replacement humanoid. GMO says it is the first maintenance van in Japan built specifically for that job. Service is slated to start in Tokyo later this month from a Shibuya lab, with a pitch of reaching central sites in one to two hours. There is one vehicle for now.

Inside, the joke becomes a workshop. Japanese press photographed a stretcher sized for a humanoid, diagnostic kit, consumables and room to lay a machine down. If the crew can fix the unit on the floor, they do. If they cannot, the broken robot rides back to the lab and the spare stays on the job. A monthly subscription already on GMO’s books — reported around ¥70,000 — is how some customers get that swap. President Tomohiro Uchida framed the van as infrastructure for social implementation, not a stunt. That is the right category even if the paint job is doing comedy.

The company does not build the robots. GMO AIR is a Unitree sales agent in Japan. The fleet it is promising to keep alive is largely G1-class hardware already in demos, including a Japan Airlines luggage trial at Haneda that started in May. The ambulance idea itself, engineers told Japanese media, came from a World Humanoid Robot Games heat: a GMO-entered Unitree went down in a 400-metre race, the crew swapped a head in about ten minutes, and the machine made the next round. One researcher put a rough number on wear — after three months, perhaps 30 to 40 percent of units see some part fail. That is an engineer talking, not a published reliability study.

Read as operations, the van is the missing sentence in every humanoid pilot. A biped that cannot be fixed before the next shift is a sculpture. Roadside assistance is how cars became infrastructure. GMO is trying the same move with a purple light bar. It is still a human-driven van. Nobody should write that an autonomous ambulance is coming to rescue robots. The intelligence, such as it is, sits in the spare unit and the person with the laptop.

Scale is the open question. One truck covering Tokyo is a press event. Years-out talk of regional bases is a plan. Japan’s broader robot targets — millions of machines across sectors by 2040, in the national paperwork — only work if someone owns downtime. GMO also launched two companion services the same day: field data fed back into performance, and network monitoring against tampering. Those are the unglamorous half of implementation. The van is what photographs well.

For this page the story is not the livery. It is that a reseller who already has humanoids in an airport decided the next product was a recovery vehicle. If the Shibuya crew logs real callouts and posts time-to-restore, the ambulance becomes a small piece of industry. If it stays a wrapped van in Minato, it is a good clip. Either way, every vendor promising a humanoid on a site now has a question they cannot dodge: when it falls over, who shows up, and what do they bring.

A camera crew walked Knightscope’s Silicon Valley headquarters and let founder William Santana Li talk through the machines. The company, founded in 2013 and listed on Nasdaq as KSCP, builds autonomous security robots, emergency call boxes and the software that ties them together. The figures it is repeating this month match the visit: 434 clients, 42 U.S. states, more than 4.4 million autonomous hours logged. Those numbers are in the company’s own Q2 materials and in the briefs it has been issuing around GSX in Atlanta this week. They are company-reported. They are also consistent across filings and press rooms, which is more than most robot startups bother with.

On the floor the lineup is wheeled, not bipedal. The K5 is the familiar cone-shaped patrol unit. The K7 is the newer, larger outdoor machine aimed at yards, hubs and infrastructure; Knightscope has said first deployments are due in the fourth quarter of 2026 after an alpha-to-beta gate. Around them sit K1 towers and capsules, emergency comms, and a prototype “H1” augmented agent. Li’s pitch is not that a robot replaces a guard. It is that cameras see everything and decide nothing, guards cannot be everywhere, and a managed “Autonomous Security Force” — machines, software, licensed people under one contract — is supposed to close the gap.

That Force is the story Knightscope is selling at GSX 2026 in Atlanta through 16 September: an ASF-7 escalation ladder with a human still holding authority at every level. The HQ video is the factory-floor version of the same argument. You see the hulls, the sensors, the command talk. You do not see a humanoid. Knightscope is adjacent to this page because it is physical autonomy in public space with real hours, not because it walks on two legs. Saying that out loud keeps the brief honest.

The financials behind the tour are mixed in the usual public-company way. Q2 2026 revenue hit a record $9.0 million, up sharply from a year earlier. Gross margin turned slightly positive. Operating costs and net loss also jumped as the firm hired and bought its way into a fuller security offering. Cash at mid-year was thin relative to the ambition. Recurring revenue is most of the mix. Li talks about a $230 billion U.S. security TAM and an in-account runway he says Knightscope has barely touched. Investors can read the 10-Q. This brief only needs the operational claim: the robots are already on sites, in volume that a humanoid vendor would envy.

What Li is willing to say on camera is where he thinks the industry goes: fewer stacked vendors, one accountable watch, machines on the monotony, people on judgment. The machines will keep logging hours whether or not a humanoid ever joins the parking lot. If a biped vendor wants those hours, it will have to answer the same questions Knightscope already lives with — who is on the radio at 3 a.m., who owns the incident report, and what happens when the hull takes a hit.

For a humanoid daily, the useful takeaway is infrastructure envy. Four million-plus autonomous hours is what a security robot company looks like after thirteen years of unglamorous patrol. GMO’s new ambulance is the humanoid market admitting it does not have that yet. NEURA’s IFA speech is Europe saying it wants an ecosystem before the hours exist. Knightscope already has the hours and still is not a humanoid firm. That contrast is the news, not the livery on a K5.

Microduck is the second consumer robot from Pollen Robotics, the Bordeaux team now inside Hugging Face, and it is built to waddle across a desk rather than stand on it. Official specs put the biped at 25 centimeters tall and about 780 grams, with 15 degrees of freedom through the legs, neck, head and an articulated beak. Out of the box it walks, sits, crouches, roller-skates, picks things up and gets back on its feet after a fall. You drive it with the included gamepad. Pre-orders opened late August at $399 before tax and shipping, in Cream, Graphite, Lavender and Sky. First deliveries are aimed at Christmas 2026 in North America and Europe.

The sensor list is the reason researchers are paying attention. A front camera, an 8×8 time-of-flight LiDAR, two IMUs, microphones, a speaker, NFC in the head and beak, Wi-Fi and Bluetooth sit on a Rockchip RK3566 with a small AI accelerator, 1 GB of RAM and 32 GB of storage. Battery is a removable NP-F550 good for about an hour. Software — SDK, simulation and the reinforcement-learning stack — is open source. The mechanical and board files are not. Pollen is explicit about that distinction, and it should stay in the copy: this is an open training platform, not open hardware.

That is the product thesis. Reachy Mini was a talking head for interaction. Microduck is a body you can train. Seven behaviors ship loaded. New ones are meant to be learned in simulation and dropped onto the real duck. Hugging Face has already talked up early pre-order volume in the five figures; later orders have been slipping past the Christmas window. A $399 biped that falls down and stands up is catnip. Whether the sim-to-real loop is pleasant on a machine with 1 GB of RAM is the part nobody can review until boxes land.

Treat it as a research toy with a storefront, not a household appliance. Accessories — roller feet, spare batteries, a dock, NFC-tagged objects — are how Pollen expects people to keep playing after the first afternoon. Races and football with several ducks are in the pitch. The honest sentence is still the short one on the product page: a small biped you can drive today and teach tomorrow, if the shipping date holds.

The storefront has already had to say that out loud. A banner on Pollen’s checkout warns that the community ordered a lot of ducks, Christmas is no longer promised on new baskets, and fresh orders are looking at four to six months. That is the usual tax on a viral pre-order, not a verdict on the robot. What matters for this page is narrower: Microduck is a biped you can train, priced like a gadget, shipping first to people who clicked in August. Everyone else is buying a place in line.

DaxAI Robotics brought a saddle and handlebars to the quadruped and called it a horse. The Qiji X1 is a full-size, four-legged mount that a person can sit on, and it spent August drawing queues at the World Robot Conference in Beijing. Riders crossed the carpet. At least one humanoid climbed on as a live payload. The company, a young Beijing startup, also showed the machine earlier in the summer at WAIC in Shanghai. “World’s first” is the marketing line. What is verified is simpler: a working, rideable quadruped on a show floor, not a render.

Company figures circulating with the debut put payload around 300 kilograms, top speed around 7 to 10 km/h on the pure-legged X1, and range near 40 kilometers on a high-voltage pack. The X1 itself is in the 300-kilogram class. A wheeled-leg XS variant was shown alongside it. Chinese coverage has put the base X1 near 289,000 yuan, with a three-year JD.com tie-up and pre-sale talk attached to the WRC stand. Those are vendor numbers. There is still no independent off-road test or published safety case.

The pitch is terrain wheels hate: slope, gravel, mud, a private estate, a forest path. Electric joint motors do the walking; onboard models are supposed to pick footfall. In the hall the gait looked deliberate and a little unsteady, which is what you expect when a 300-kilo machine is learning to carry a tourist. Control from the bars is the story DaxAI wants. Some of the stage passes were still being flown from the side with remotes. Both can be true at an expo.

This is a spectacle product with a logistics aftertaste. If the payload holds up off the carpet, you have a slow pack animal that does not need a trail graded for tires. If it does not, you have a conference ride. Until someone publishes hours, fault rates and what happens when a rider comes off, the Qiji X1 stays where it is: a large quadruped people actually sat on in Beijing, and a reminder that Chinese labs will try the joke form factor until it stops being a joke.

It is also only loosely a humanoid story. The reason it sits on this page is the rider — a person, and once a humanoid, using the same hall that is otherwise full of bipeds. DaxAI is selling a mount, not a worker. Limited deliveries have been talked about for late 2026. Until a unit leaves the expo circuit and comes back with a scratched belly and a log file, treat the horse as a show that worked, not a category that has arrived.

The 2026 China Mining Conference and Exhibition opened in Tianjin on September 10 at the Meijiang Convention and Exhibition Center, the 28th edition of a show the organizers bill as one of the four major international mining gatherings. This year’s theme is win-win cooperation with a green and intelligent turn. More than 650 companies are on the floor through September 12, and official photography already treats robots as part of the scenery: a mine-inspection robot dog on a stand, unmanned equipment in the aisles, ministerial forums on international mining cooperation in the halls. The exhibition is the place China uses to argue that a historically dangerous, labor-heavy industry is being rebuilt as a high-tech system. The argument is no longer abstract. Quadrupeds, cabless electric haul trucks and underground networks are being shown as products that already have sites, not only as concept boards.

One verified floor exhibit is a four-legged inspection robot built to enter hazardous, enclosed underground spaces where sending a person is the last option. Xinhua recorded a quadruped patrol robot on a preset underground route, adjusting for obstacles and working in closed headings; the named booth in that official dispatch was Ziguang Tianji. The job is not glamorous. A dog-sized machine walks a heading, carries cameras and gas or thermal payloads, and sends a picture back before a crew steps in. For non-coal metal mines and for coal sites under tighter safety rules, that first look is the difference between a planned inspection and a person in an unknown atmosphere. Tianjin is selling that substitution as routine industrial kit. Vendor names that appear only in secondary video narration, and that could not be matched to an exhibitor list or official photo caption, are not used here.

Haulage is the other half of the pitch. Exhibitors are showing unmanned, purely electric mining trucks that move ore without a driver in the cab. That class of vehicle is already in commercial use in China, separate from any single booth claim. Official and industry tallies put the in-service unmanned mining-truck fleet above 4,000 units by the end of 2025, with roughly a tenth of the country’s mining trucks now driverless. Large open-pit sites such as Yimin in Inner Mongolia have been running cabless electric fleets under 5G-Advanced networks. The commercial point at Tianjin is not that a truck can drive itself in a video. It is that the truck, the charger, the dispatch cloud and the safety case are being sold together so a mine can take people off the haul road in fog, snow and dust — the conditions that used to stop work because a human driver could not see.

The connective tissue under both the dog and the truck is communications and software. Chinese operators have been installing underground low-latency 5G and building 3D digital replicas of workings so a control room can watch a live twin of the heading rather than wait for a shift report. Digital twins of hoists, tailings levels, equipment load and personnel location were on display in smart-mine sand tables at the same show. The combination is what vendors mean by “intelligent mining”: the robot or the truck is only the visible actuator. The network keeps latency low enough for remote or supervised autonomy, and the twin is the place supervisors look when they decide whether a heading is safe to enter.

None of this makes a mine empty of people. What it changes is which jobs remain underground. Fixed posts and the first pass through a bad heading are the tasks regulators and operators most want off human backs. China’s mine-safety authorities have already pushed pilot programs for robots in dirty, dangerous and exhausting roles. Tianjin is where that policy meets a sales floor. A quadruped in a booth and a cabless truck on a screen are exhibits. A heading that no longer needs a walker with a lamp is the deployment. The gap between those two is still the story: certification, explosion-proofing, battery life in heat and dust, and whether the twin on the wall matches the ground when the network drops.

For the robotics market the mining floor is a reminder that the paying customer is often not a humanoid buyer. Inspection dogs, rail-mounted patrol robots, unmanned trucks and remote shovels are absorbing the same sensors, radios and autonomy stacks that humanoid firms advertise for factories. If China can keep scaling unmanned haulage and underground patrol, the data set that trains the next inspection model will come from mines, not from living-room demos. Tianjin this week is that pipeline in public: a national mining show using robots to say the industry’s next productivity gain is fewer people in the heading, not only bigger shovels at the face.

XPENG has switched on what it calls the world’s first automated production line for an advanced general-purpose humanoid, and it marked the moment with a robot that finished final assembly and walked off the line without a handler on the tether. The commissioning, announced in Guangzhou in the second week of September 2026, is the company’s bid to move IRON out of R&D prototyping and onto a factory schedule. XPENG says more than 80 percent of the line’s core processes are automated and that the process control is borrowed from its electric-vehicle quality system. That claim is the whole story in miniature. A humanoid that can take a step after the last station is a manufacturing object. A humanoid that still needs a puppeteer to leave the cell is still a prototype, however polished the walk looks on stage.

IRON is the same platform that went viral in November 2025 when XPENG cut a leg open on stage to prove no one was inside. The production description now circulating with the line is specific: 76 degrees of freedom across the body, 21 in each hand, a fully enclosed flexible lattice shell, and three in-house Turing AI chips rated at up to 2,250 TOPS so the company’s physical-AI model can run on the robot. XPENG’s point in repeating those numbers next to a factory clip is that the unit leaving the line is supposed to be the same architecture it intends to build in volume, not a one-off show machine. Shared parts with the EV business are part of the same argument. The company has said more than 85 percent of IRON’s supply chain overlaps the car side, which is how an automaker tries to make a humanoid line look financeable.

The schedule attached to the line is still a target, not a shipment log. XPENG is holding to mass production by the end of 2026, first deployments in its own stores and campuses, and a broader commercial launch with deliveries in China and overseas in 2027. Earlier in the year the firm talked about capacity above 1,000 robots a month and a longer ambition toward much larger annual volume by 2030. Tuesday’s line announcement did not restate every one of those figures, and XPENG has not published how many units the first small-batch run will yield or what a robot will cost. That omission should stay visible. Starting a line is not the same as filling a yard.

What walked off the line is being sent first into XPENG’s own world. Showrooms and campuses are the proving ground: tours, reception, and other supervised public tasks that generate data without asking a third-party factory to bet a shift on an unproven machine. The company has said additional in-house scenarios will be verified in the fourth quarter of 2026 before it tries a wider commercial push. That sequence is the automotive habit applied to a humanoid — build it, walk it, use it internally, then sell it. Competitors will watch whether the walk off the line is repeatable at the next ten units or only at the camera unit.

The “world’s first” language is XPENG’s own and should be read as a category claim about a high-DOF general-purpose humanoid leaving an automated line under onboard control. Other firms have built humanoids and other firms have built robot factories. What XPENG is advertising is the combination: automotive-grade automation, a production IRON, and an unaided walk at the end of the cell. Independent coverage has treated the clip as the clearest sign yet that IRON is heading for manufacturing rather than another stage demo. It has also noted the comparison hanging over the news cycle: Tesla’s Optimus program has not shown an equivalent line moment. That comparison is market color. It is not a verdict on either robot’s field reliability.

The industrial test from here is ordinary. Yield at the end of the line, hours before the first unplanned stop, and whether a store-floor IRON can do useful work without a safety driver in the wings will decide if Guangzhou started a product or staged a ceremony. XPENG has the capital, the EV supplier base and now a line that a robot can leave on its feet. Mass production by year-end is the date the company has put on the wall. Until units show up in numbered batches, the accurate sentence is the one the factory clip actually supports: an advanced general-purpose humanoid has completed automated production and walked off the line by itself.

AgiBot Research has released GE-Act 2.0, a world-action model trained from scratch on embodied manipulation data rather than inherited from a general video generator. Visual representation, future-state generation and action prediction are learned from random initialization. The September 2026 paper argues that most world-action models still borrow pretrained video backbones and never cleanly test whether manipulation-specific pretraining scales. GE-Act 2.0 is the counter-experiment: no inherited video generators, no task-specific fine-tuning at evaluation, and a real-robot zero-shot protocol on unseen scenes, 100 atomic tasks, 20 skill categories and two embodiments.

The architecture keeps planning and control aligned while allowing the pieces to pretrain on different data. A control-oriented autoencoder compresses each frame into tokens meant to retain instruction- and action-relevant detail. A single-step visual planner predicts a complete future state in one pass, and an inverse-dynamics model turns that future into actions. The modules are jointly trained with knowledge-aligned selective optimization, which keeps only predicted futures judged compatible with the recorded action. The project page says a full action chunk runs in 104 milliseconds on a single RTX 5090, or 52 executable actions at 30 hertz.

The scaling curve is the result the field will cite. Co-training data expanded in nested pools from 300 hours to 1,200, 5,000 and 30,000 hours. On G1-OP, mean zero-shot success rose from 17.1 percent to 44.1 percent. On G2-90D it rose from 13.4 percent to 31.1 percent. Gains covered 19 of 20 skill groups on G1-OP and 18 of 20 on G2-90D. Tasks with any nonzero success increased from 39 to 76 on G1-OP and from 24 to 72 on G2-90D, and the curve did not flatten between 5,000 and 30,000 hours.

New skills appearing only at the largest scale matter more than the averages. Folding towels, nesting paper cups, uncapping pens and arranging flowers are the examples AgiBot emphasizes — manipulations the 300-hour model could not complete. Skill-specific data coverage correlates strongly with zero-shot out-of-distribution success, with Pearson r of 0.80 and Spearman rho of 0.85. Under the same protocol the model grounds object, color, shape and position references in at least 90 percent of trials and follows explicit instructions even when they conflict with an already-committed behavior.

Cross-embodiment transfer is the second claim that will be tested. G2-90D accounts for less than 2 percent of the co-training mixture, yet still gains 17.7 percentage points as the shared corpus grows. G1-OP supplies more than half of the co-training data, so the experiment tests whether a data-rich body can lift a data-poor one. The mixture includes instruction-video data, action-labeled trajectories, simulation, open-source sets, egocentric human video, deployment rollouts and failure trajectories. Failure is treated as a training asset rather than discarded noise.

The industrial reading is unfinished. A model that envisions a future state before it acts is closer to how world-action systems were supposed to work than a policy that only clones teleoperated trajectories. GE-Act 2.0 is still a research checkpoint, and 44.1 percent zero-shot success on a 100-task wall is not factory reliability. What the result changes is the burden of proof for teams that fine-tune a web video model and call the leftover gap “embodiment.” If a from-scratch model keeps climbing from 5,000 to 30,000 hours, the next argument is how expensive the next tenfold data increase will be.

A special letter from the desert is how DEEP Robotics is framing its latest field deployment. This summer the Hangzhou company accepted an invitation from the Alxa League oasis plantation center and sent a Lynx wheeled-legged robot into the Tengger Desert to support patrol and sand-control work. The managed belt includes more than 100 mu of flower-rod forest and some 40,000 sand date trees, a thin green edge that rangers have historically walked 15 to 20 kilometers a day to inspect. DEEP presents the robot as a force multiplier for people already doing the work rather than a replacement for the ecological program.

The machine in the footage is the Lynx wheel-legged platform, described here as the M20S variant. Unlike a conventional quadruped that only walks, Lynx can roll when the crust will support speed and articulate the legs when the surface turns to soft sand, ruts or broken fencing. Company materials have cited a top speed up to 9 meters per second on the M20S and a payload that in the Alxa write-up reaches 35 kilograms of tree tags, water bags and replanting tools. That load is the core of the job: fewer return trips across hot sand for the same set of tools.

The environment is the actual product test. Surface temperatures around 50 degrees Celsius, wind-driven grit and long unstructured routes are closer to industrial inspection than to a motion-capture stage. The Alxa account says the robot held a patrol line across loose sand, used a dual-light pan-tilt to flag holes, downed plants and torn nets, and relied on IP67 sealing against dust. A thermal payload tracks wildlife near young plantings, and a loudspeaker is used to drive animals off seedlings without harming them. The claim is that 7-by-24 patrol coverage becomes possible where human routes are long and visibility collapses after dark.

This desert assignment sits on a longer DEEP field record. Lynx M20 was introduced in 2025 as an industrial mid-size wheel-legged robot for tunnels, substations, emergency logistics and scientific support, with operating bands that typically run from minus 20 to 55 degrees Celsius and protection ratings of IP66 or IP67 depending on the variant. The same family has been shown on red-sand slopes, in Turpan vineyards moving harvest baskets, and as a logistics mule on high-plateau research trips. The Tengger deployment is an application layer on a mobility stack already sold into heat, dust and payload work.

What the video is useful for should be kept separate from what it is not. It is useful as evidence that a wheel-legged robot can stay upright on terrain that breaks ordinary wheeled carts and exhausts walking patrols. It is not a controlled study of hectares restored, seedling survival or cost per inspected kilometer. Desertification control is a multi-year land-management problem. A robot that carries water bags and deters grazing animals can change the labor equation on the margin; it cannot substitute for planting density, irrigation and water politics in Alxa.

For the wider market the story is a reminder that not every commercially interesting robot this year is a biped in a factory. Extreme-terrain logistics, ecological patrol and inspection in places people should not walk twice a day are already paying use cases for Chinese quadruped and wheel-legged vendors. If Lynx can keep a schedule in the Tengger — heat, payload, dust, night thermal and a real ranger workflow — that is a more transferable data point than another indoor climb. The open test is durability across a full planting season, not a single summer visit.

Chinese EV maker XPeng has officially opened its humanoid robot assembly line, marking a significant milestone in the transition from research and prototyping to full-scale production. The company’s IRON platform has completed automated production and autonomously walks off the line, signaling the start of a new era for advanced general-purpose humanoids in China. Facing a slowdown in electric vehicle sales, XPeng is now leveraging its manufacturing expertise and supply chain to scale humanoid output, with the goal of reaching meaningful volume production by the end of 2026.

This move comes at a critical time for the industry, as several Chinese automakers are expanding into robotics to diversify revenue and capitalize on the growing humanoid market. XPeng’s IRON robot is now walking off the assembly line autonomously, a first for the sector and a demonstration of the company’s ability to integrate software and hardware at scale.

While China dominates approximately 97 percent of humanoid robot exports, other players are also active. Figure AI’s Figure 03 continues to perform paid logistics work at BMW facilities, and Boston Dynamics is shipping fleets of its Electric Atlas robots to Hyundai factories in South Korea. These deployments highlight the global reach of humanoid technology beyond China’s borders.

The timing of XPeng’s production ramp-up is notable amid tightening export controls and supply-chain scrutiny. The company’s ability to execute on its volume targets will serve as a benchmark for whether Chinese automakers can deliver competitive humanoid platforms that match Western competitors on both technical and economic grounds.

Challenges remain in scaling from low-volume pilots to true mass production, including software optimization, supply-chain integration, and rapid iteration based on real-world feedback. XPeng’s success will be closely watched as it helps shape the overall industry timeline for 2026 and beyond.

The line itself achieves more than 80 percent automation across core processes, with robots handling the majority of assembly tasks and the completed IRON unit walking off under its own power in a fully autonomous workflow. This level of integration demonstrates XPeng’s ability to combine advanced factory robotics with humanoid-specific software, setting the stage for consistent, high-volume output while minimizing human error and labor dependency.

Performing pirouettes, tennis strokes, or celebrating goals with the naturalness of a human is now possible for a robot. The BeyondMimic project presents a new robotic learning framework powered by human movement data, which creates natural physical behaviors. The robot can walk, run, and execute complex movements inspired by martial arts after training with approximately 2.5 hours of human movement data. The researchers first validated their method in a simulation environment before conducting physical experiments on different terrains.

This approach allows the robot to learn from real human demonstrations and translate them into smooth, human-like physical actions without requiring massive amounts of data or complex custom controllers. The framework’s ability to capture and replicate the fluidity of human motion opens new possibilities for robots in service, entertainment, and even collaborative tasks where natural interaction is essential.

Physical experiments confirmed that the trained policies generalize across varied surfaces and conditions, highlighting the robustness of the approach. This work represents a significant step toward more versatile and responsive robotic systems that can move and act in ways that feel intuitive to humans.

The BeyondMimic framework is positioned as a key advancement in embodied AI, with potential applications ranging from everyday household assistance to professional sports or performance environments. As the field moves forward, continued improvements in data efficiency and real-world transfer will determine how quickly human-like movement becomes a standard feature of advanced robots.

By leveraging open-source motion-capture tools and a two-stage pipeline—first mastering motion tracking via reinforcement learning, then enabling versatile control through guided diffusion and classifier techniques—BeyondMimic bridges the gap between scripted tasks and fully adaptable behaviors. This efficiency is particularly promising for rapid deployment in dynamic environments where robots must respond intuitively to unstructured real-world conditions.

The researchers note that while action scripts are still required for high-level task planning in testing phases, the core movement policies learned through BeyondMimic require minimal manual intervention, accelerating progress toward fully autonomous humanoid systems that can seamlessly integrate into collaborative or entertainment settings.

Real innovation cannot be built on slides. From hardware to AI, true breakthroughs demand persistence and a willingness to get back up after every fall. XPENG Robotics has walked side by side with He Xiaopeng for six years, pushing forward through setbacks to build robots that interact naturally, hold memories, and deliver real-world value. Today, with the company’s production line now fully active and its humanoid platform advancing, He Xiaopeng continues to emphasize that we take it one step at a time—no shortcuts, no quick fixes, but steady progress toward machines that can truly operate in human environments.

This philosophy has defined XPENG Robotics’ journey from laboratory experiments to full-scale manufacturing. He Xiaopeng, who also serves as CEO of the newly formed XPENG-W Robotics unit, has personally guided every milestone, ensuring that technological progress aligns with practical deployment realities. The company’s focus on embodied intelligence means that every robot must be reliable, adaptable, and ready for real environments, not just controlled simulations or idealized demos.

Be patient—we take it one step at a time. This mantra reflects the deliberate pace at which XPENG is scaling its humanoid platform. From initial prototypes to production-ready units, the emphasis remains on quality, iteration, and real-world testing rather than rushing timelines or cutting corners. The result is platforms that can hold memories, respond naturally to human interaction, and create tangible value in homes and workplaces, exactly as the six-year partnership has aimed to deliver.

Looking back over six years, the persistence has paid off in multiple ways. XPENG’s transition from autonomous vehicles to robotics demonstrates the company’s ability to leverage its automotive expertise in building intelligent machines capable of complex physical tasks. He Xiaopeng’s leadership ensures that every decision—from hardware design to AI training—prioritizes long-term capability over short-term gains, turning challenges into opportunities for deeper understanding and better results.

As the production line ramps up and the first humanoid units move toward deployment, the industry watches closely. Will XPENG deliver on its promise of natural interaction and memory retention at scale? The answer will shape how quickly humanoids move from research demos to everyday companions that hold memories and create real-world value. For now, the message remains clear: we build it right, one step at a time, with He Xiaopeng at the center of every advance.

A video from Russia showing a humanoid robot striking a customer inside an electronics store has sparked a fresh round of debate about what happens when humanoid machines move out of controlled demonstrations and into everyday public spaces. In the footage, the robot repeatedly makes contact with a man standing nearby, using both its arms and legs before the encounter eventually ends.

The scene is unusual precisely because it looks so ordinary. There is no laboratory setup or carefully choreographed product demonstration—just a customer standing in a retail environment when a machine suddenly becomes physical. That makes the footage particularly relevant to the broader robotics industry, where humanoids are increasingly being tested in stores, factories, warehouses, and other places where people and machines have to share the same space.

The exact circumstances behind the incident remain unclear. The footage alone does not establish whether the behavior resulted from a software malfunction, a control error, an unexpected interaction, or some other failure. That distinction matters, because a humanoid robot capable of walking, manipulating objects, and responding to people also needs to recognize when physical contact is unsafe and immediately transition into a controlled state.

That safety layer is becoming one of the most important challenges in consumer robotics. A robot working around people cannot rely only on its ability to complete a task; it also needs reliable perception, force control, collision detection, emergency-stop systems, and clearly defined operating boundaries. In a retail environment, those requirements become even more important because customers are not trained operators and may not know how a machine will react to an unexpected movement.

The incident also highlights a larger gap between impressive demonstrations and dependable real-world deployment. Humanoid robots are designed to operate in environments built for humans, which means they inevitably encounter unpredictable situations: people stepping into their path, objects being moved unexpectedly, children approaching them, or customers interacting with them without understanding their limitations. Handling those situations safely is just as important as demonstrating dexterity or conversational intelligence.

For developers, the lesson is less about one unusual video and more about the standards required as humanoids become physically present in public life. Every deployment creates another opportunity to test how a robot behaves when conditions differ from the assumptions made during development. Transparent reporting, stronger safeguards, and well-defined emergency procedures will become increasingly important as companies move from pilot programs toward larger commercial deployments.

No injuries have been reported in connection with the footage, and the available video does not by itself establish the underlying cause of the behavior. But the episode is a useful reminder that humanoid robotics is ultimately a physical discipline. A system can be remarkably capable in software and still present serious challenges if its physical behavior is not predictable, controllable, and safe around ordinary people.

XPeng has activated its dedicated humanoid production line, marking a significant milestone in the company’s transition from autonomous vehicles to robotics. The facility is now running at full capacity, with the company aiming for meaningful volume output by the end of 2026 as part of its broader strategy to scale humanoid production alongside its existing EV and smart-mobility businesses.

The move comes at a time when several Chinese automakers are expanding into robotics to diversify revenue streams and leverage their existing supply-chain and manufacturing expertise. XPeng’s IRON humanoid platform is already in advanced development, and the new line is designed to support both initial pilot deployments and eventual mass production. The company has emphasized that the production ramp will focus on reliability and cost control while maintaining the high standards required for consumer and industrial use.

Industry analysts view the XPeng production line as another data point in the growing wave of Chinese humanoid initiatives. While Western companies have focused primarily on research and limited pilots, several domestic players are now moving into actual manufacturing. The timing is notable because it coincides with tightening export controls and supply-chain scrutiny, making domestic production a strategic priority for both cost and security reasons.

Challenges remain. Scaling from low-volume pilots to true mass production requires not only hardware but also software optimization, supply-chain integration, and the ability to iterate quickly based on real-world feedback. XPeng’s success will serve as a benchmark for whether Chinese automakers can deliver humanoid platforms that are competitive on both technical and economic grounds.

The broader implication is that 2026 is shaping up to be a pivotal year for humanoid volume production. Multiple Chinese companies are now in or near production, which could accelerate the overall industry timeline and make early deployment in factories, logistics, and consumer markets more likely than in previous years.

For investors and competitors, the XPeng production line represents both opportunity and urgency. The company’s ability to execute on the volume target will determine how quickly the humanoid market can move beyond demos and into tangible economic impact. Whether XPeng’s output is sufficient to meet its ambitious timeline will be one of the key watchpoints for the remainder of 2026.

Microduck is a compact, open-source humanoid robot designed for education, research, and playful experimentation. At just 25 cm tall and weighing 780 g, the 15-degree-of-freedom machine is small enough to fit on a desk yet powerful enough to perform a wide range of autonomous and interactive behaviors right out of the box.

The robot features a front camera, 8x8 LiDAR, dual IMUs, microphones, speaker, NFC reader, Wi-Fi, and Bluetooth. Out of the box it can walk, sit, crouch, roller-skate, and recover from falls on its own. It can also pick up objects with its articulated beak and learn new tricks through reinforcement learning or manual control. The software stack is fully open-source, allowing developers to extend behaviors, add new sensors, or even train custom policies.

Pre-orders are now open at $399 (before taxes and shipping), with first deliveries targeted before Christmas 2026. The robot is available in four color options: Cream, Graphite, Lavender, and Sky. The low price and open-source nature make it accessible to hobbyists, educators, and researchers who want to experiment with embodied AI without the usual high cost of commercial platforms.

Microduck sits at the intersection of consumer electronics and robotics. It is not intended to replace industrial humanoids but rather to bring the joy and learning of robotics to a much larger audience. The open-source approach also accelerates innovation because the community can contribute improvements, new behaviors, and even hardware designs.

While the robot is small, its capabilities are surprisingly capable for its size. The combination of reliable self-recovery, real-time perception, and modular expansion makes it a versatile platform for teaching, research, and creative projects. Whether you are a parent introducing robotics to a child, a researcher testing new control algorithms, or an enthusiast building a custom swarm, Microduck offers a practical entry point into the world of embodied AI.

The pre-order window is open now, and the company has already started taking deposits. The tight shipping schedule before Christmas suggests the first units will ship in late November or early December, giving early buyers a chance to test the robot over the holiday season and provide feedback that will shape future versions.

AGIBOT has demonstrated a striking example of coordinated multi-robot performance with its A3 humanoid platform, showing three synchronized robots executing a fluid dance routine. The short video captures the robots moving in perfect rhythm with one another — a clear illustration of the coordination capabilities that AGIBOT is developing for both entertainment and industrial applications.

The demonstration highlights the company’s progress in embodied AI and multi-agent systems. By controlling multiple A3 units as a single team, the robots achieve smooth synchronization that would be impossible with independent single-unit deployments. This capability directly advances AGIBOT’s broader goal of deploying fleets of humanoids that can collaborate in real-world settings such as factories, logistics hubs, or entertainment venues.

AGIBOT’s platform continues to evolve rapidly. The company is not only advancing its A3 humanoid but also scaling industrial deployments, exploring embodied AI applications, and preparing variants such as the A3 Ultra with enhanced autonomy and manipulation. The synchronized dance performance is presented as both a fun showcase and a technical milestone that underscores the company’s ability to move from individual robot capabilities to multi-robot teamwork.

While the video itself is entertaining, the underlying message is technical: AGIBOT is pushing the boundaries of robot coordination and control stacks. The success of the 3-robot routine relies on advanced perception, predictive control, and robust communication between units — areas where the company is investing heavily to prepare humanoids for collaborative industrial and service roles.

AGIBOT’s broader platform provides context: the company is not only developing the A3 humanoid but also running industrial deployments, exploring embodied AI, and scaling its A3 Ultra variant with improved autonomy and manipulation. The synchronized dance is therefore presented as both a showcase of athletic capability and a cautionary tale about the engineering effort required to make such behaviors repeatable in the physical world.

Ultimately the synchronized dance performance is less about the choreography itself and more about the ongoing challenge of closing the gap between impressive demonstrations and reliable, scalable multi-robot deployments. For the wider sector it serves as both an inspiration and a warning: impressive simulation results are necessary but never sufficient. The real test of any humanoid platform is how well its learned behaviors survive the transition to unpredictable physical conditions, and AGIBOT is using its own platform to illustrate exactly where that transition still breaks.

Chinese startup DaxAI Robotics has unveiled the Qiji X1, a full-size four-legged robot platform designed so a human can ride it like a horse while the built-in AI handles navigation, safety, and basic task assistance. Debuted at the 2026 World Robotics Conference in Beijing, the Qiji X1 represents a new category of transport robot that blends mobility, autonomy, and human-rider interaction in a single package.

At its core is a rideable chassis with four powerful legs that can traverse rough terrain, stairs, and uneven surfaces where traditional wheeled or tracked vehicles would struggle. The AI system, developed in-house by DaxAI, includes real-time perception, path planning, and emergency-response capabilities, allowing the rider to focus on steering while the robot manages balance, obstacle avoidance, and basic environmental awareness. The platform is positioned as both a personal transport device and a potential base for future delivery or inspection tasks.

Pricing has been reported around $40,000 for the initial versions, placing it in a premium segment that combines the cost of a high-end vehicle with the capability of a specialized robot. The debut at the World Robotics Conference underscores DaxAI’s ambition to bring fully autonomous, rider-capable machines from research labs into practical use. The company has framed the Qiji X1 as the first true “robot horse” in the sense of a four-legged platform where a human can ride and control it directly.

While the platform is still in early deployment, its integration of AI, vehicle dynamics, and human-rider interface raises interesting questions about the future of personal transport. Unlike traditional electric vehicles or bicycles, the Qiji X1 can adapt its gait, recover from falls, and operate in environments that would be inaccessible to most wheeled or tracked alternatives. Whether the AI layer proves reliable enough for daily use or whether regulatory and safety concerns slow adoption will shape the product’s trajectory.

DaxAI’s announcement also fits into a broader trend of Chinese companies exploring rideable or passenger-carrying robots. The Qiji X1 demonstrates that the hardware and software foundations are now mature enough to consider human riders in real-world applications, a milestone that moves beyond purely industrial or research use cases. For the sector it serves as a reminder that the ultimate value of humanoid and quadruped platforms may lie not only in labor replacement but in new forms of mobility and assistance that combine mechanical strength with intelligent decision-making.

The Qiji X1 is therefore best read as an early but concrete example of how rideable robot platforms could evolve from niche concepts into practical consumer and service vehicles. Whether it becomes a bestseller in China or remains a conference demo will depend on real-world reliability, regulatory approval, and the long-term cost of operation. For now, the debut at the 2026 World Robotics Conference has successfully shown that a human can ride a fully autonomous four-legged robot horse—and that the technology is ready for the next stage of development.

XPENG has closed its latest financing round for its dedicated robotics business, securing more than $900 million in fresh capital at a valuation that places the unit above $6.3 billion. The round, led by IDG Capital and backed by established Chinese investors, marks a significant step toward the mass-production target for the company’s IRON humanoid platform, with initial deployments now projected for late 2026.

The funding round comes at a moment when several Chinese automakers are accelerating their own robotics programs. XPENG has positioned the IRON robot as a versatile platform capable of supporting both factory and service-sector applications, and the valuation reflects growing market expectations that domestic players can maintain technological momentum even under export restrictions and heightened scrutiny.

Industry commentary has focused on the timing and size of the raise. Wall Street analysts have noted that the $6.3 billion figure sets a new benchmark for Chinese humanoid developers and may influence future valuation discussions for other domestic platforms. The capital will be used to accelerate production lines, expand sensor and AI capability, and begin pilot deployments with strategic partners in manufacturing and logistics.

Challenges remain. Delivering on the late-2026 production target requires not only hardware scaling but also software maturity, supply-chain reliability, and the ability to iterate quickly on real-world feedback. XPENG’s success will serve as a data point for whether Chinese companies can close the gap on Western programs while operating in a more fragmented global environment.

The round also highlights a broader trend: Chinese automakers are treating robotics as a strategic growth vector rather than a side project. Whether the capital flows translate into machines that actually outperform imported alternatives in cost, reliability, or regulatory approval will determine how much the sector can grow outside Western markets.

For investors and competitors alike, the XPENG round provides a clear signal that the race to produce humanoids at scale is no longer confined to a handful of Western names. The valuation and funding volume suggest that Chinese players are prepared to invest heavily and move fast, even if the resulting platforms must navigate a more complex international landscape.

Barclays has published a forward-looking assessment of global humanoid-robot adoption that projects more than 60,000 new units entering service in 2026 alone. The forecast, drawn from a combination of industry data and proprietary modeling, places the steepest growth curve in the latter half of the decade as hardware costs decline, reliability improves, and early adopter industries begin to demonstrate measurable ROI.

The 2026 figure represents the first year in which total deployments are expected to cross a meaningful threshold and includes both internal corporate fleets and the first wave of external commercial sales. The analyst team notes that the bulk of 2026 activity will likely concentrate in automotive assembly, logistics, and certain service-sector applications where tasks are relatively structured and labor availability is constrained.

Looking further out, Barclays models annual deployments reaching approximately 13 million by 2035. The long-term trajectory assumes continued cost reductions—potentially to the $20,000–$30,000 range for consumer-grade models—along with incremental gains in autonomy that allow robots to operate across a wider range of unstructured environments. The projection also factors in the possibility of regulatory frameworks that accelerate or slow adoption depending on safety and data-privacy standards.

While the numbers are high-level estimates, the shape of the curve aligns with earlier commentary from other firms that have identified 2027 as a potential inflection point for external sales. The 2026 ramp is framed as the period when early machines move from pilot programs and academy-style training into genuine operational environments, generating the first wave of real-world data that will shape the next generation of platforms.

Significant execution risk remains. Many of the projected 60,000 units in 2026 are still in the final stages of development or limited pilot deployments. Any delays in software reliability, supply-chain bottlenecks, or regulatory hurdles could compress the 2026 figure and push meaningful volume growth into 2028 or later. Conversely, a breakthrough in one of the key constraints—compute, perception, or actuation—could accelerate the timeline.

For the industry the Barclays forecast serves as a benchmark rather than a guarantee. It underscores that 2026 is the year when humanoids are expected to move from novelty demonstrations to a visible part of the economic landscape. Whether the actual number lands closer to 30,000 or 100,000 will depend on execution across hardware, software, and the policy environment that shapes both supply chains and adoption.