A Humanoid Sorts Mercedes-Benz Parts at 96%, NVIDIA's CEO Resigns, OpenAI Folds Its Robot Team

14 min listen

Apptronik's Apollo humanoid, driven by Google DeepMind's Gemini Robotics 2.5, filled parts trays on Mercedes-Benz lines at 96%, with no word on how many trays or how often a person helped. NVIDIA shipped its biggest open robot model on Monday and lost Jensen Huang as CEO on Thursday, OpenAI broke up its robotics team on the strength of a study where GPT-6 cost $36 per successful task, a lookup table with no training scored 92 on the benchmark every robot paper reports, and federal regulators found most of Waymo's school-bus incidents came after the fix.

  • Humanoids
  • Industry
  • VLAs
  • Benchmarks

Google DeepMind put an Apptronik humanoid on two Mercedes-Benz production lines this week and reported that it filled parts trays correctly 96% of the time, without saying how many trays, over how long, or how often a person stepped in. The same shape repeated all week. NVIDIA shipped the largest open robot model yet announced and its own whitepaper concedes the headline figure is "high partly by construction." OpenAI broke up its robotics team on the strength of a study where its general model beat every robot-specific model and cost $36 a task. A workshop paper found that a lookup table with no training comes within six points of the best model on the benchmark nearly every robot paper reports. And three days after shipping that robot model, NVIDIA lost its chief executive.

Gemini Robotics 2.5 on a Mercedes-Benz line

DeepMind released Gemini Robotics 2.5 on 10 November, and the result worth reading came from a factory. Apptronik's Apollo 2, fitted with the 22-degree-of-freedom SharpaWave five-finger hand, did automotive kitting on lines in Kecskemét, Hungary and Sindelfingen, Germany. Kitting means picking metal brackets out of bulk bins and laying them into trays in the order the line will use them. DeepMind reports a 96% success rate.

Two models split the job. Gemini Robotics ER, the embodied reasoning model, tracks where the robot is in the workflow. Gemini Robotics 2.5 handles the arms and the fingers. Both run on cloud servers rather than on the robot, which puts a network round trip inside a line with a fixed cycle time.

Stronger evidence of movement is the per-task comparison DeepMind published against its own previous model, Gemini Robotics 2, from 30 July 2026. Across more than 200 tasks it reports success rates 30% to 100% higher and tasks 20% to 30% faster.

Task Gemini Robotics 2 Gemini Robotics 2.5
Close the lid of a toolbox 0% 93%
Pack pastries with tongs 5% 100%
Untie ribbon bows 0% 70%
Hang a cardigan 0% 62%
Fold a pair of trousers 0% 40%

Three months is a short time to go from never to usually on tasks like these, and because the baseline is Google's own prior model the comparison is at least internally consistent.

The 96% itself arrives with no denominator. DeepMind published no trial count, no duration, no definition of what counted as a failure and no intervention rate, and the people doing the same kitting job on that line are close to error-free. Carolina Parada, who leads robotics at DeepMind, said broad generalisation may come at the cost of reaching 99.9% reliability, and director Kanishka Rao called closing the last 4% a "steep road." Saying that in a launch post is to their credit, and it still leaves a reader unable to tell whether 96% was one good shift or a sustained rate.

GXO's letter of intent and AgiBot's store log

Two days later the warehouse operator GXO signed a non-binding letter of intent with Apptronik for up to 1,000 Apollos over four years. Trials start in 2027 on picking, packing and sorting, with commercial deployment in North America in late 2027 as robots-as-a-service. The rollout is conditional on jointly agreed performance, safety and operational targets, and neither company has published them.

That structure is now the industry's standard disclosure, and "up to N robots if the numbers land" commits nobody to anything a reader can check. GXO's chief executive, Patrick Kelleher, said in August that humanoids had not returned their cost and that the company was a couple of years away. Rather than revise it, GXO has lined up a second supplier alongside Agility's Digit.

AgiBot went the other way and published something a sceptic can hold. After six of its G2 humanoids ran a convenience store in Shanghai's Xuhui District for 24 hours from 10am on 2 November, the company released the task breakdown on 13 November, counting 683 customer-service interactions, 369 product pickups and deliveries and 147 coffee orders among 1,730 in total. AgiBot also says four of the six robots passed its final assessment, scored on time on duty, task completion, customer feedback and response times, while two did not meet all the criteria. As far as I can tell that is the first time a humanoid maker has said publicly that some of its robots failed a shift. There is still no intervention count and no charging data, and the write-up is sponsored content authored by AgiBot.

NVIDIA's GR00T 2

Google's model runs in the cloud. NVIDIA's new one is built to run on the robot. At the Conference on Robot Learning in Austin on 9 November, NVIDIA's GEAR lab released GR00T 2, a world-action model rather than a vision-language-action model. One diffusion transformer generates future video frames and action chunks jointly, and small per-robot heads turn the plan into motor commands. Learning to predict pixels is what teaches it how objects move under contact, which a model trained only to copy actions never learns.

NVIDIA's research page puts 402,730 hours of egocentric human video in the pretraining mix, plus 10,184 hours of handheld UMI gripper demonstrations and 23,500 hours of robot teleoperation. Its own corporate blog says 100,000 hours of egocentric video, so the two company sources disagree. GR00T 2 Turbo is a 2-billion-parameter version with a higher-compression video encoder, and NVIDIA reports it running onboard a Unitree G1 at 124 ms latency on a Jetson Thor, planning 1.6-second action chunks about once a second while a Jetson Orin NX runs whole-body control at 50 Hz. That is the first onboard humanoid latency figure a major lab has put in writing.

The numbers below are all NVIDIA's own, on a Tianji robot with two 22-degree-of-freedom Sharpa hands.

Setting Reported success
50 unseen tasks, zero-shot 40%
50 unseen tasks, one demonstration each 71%
50 unseen tasks, five demonstrations each 95%
Seating GB300 server memory modules 85%

NVIDIA led with the 95%, and R&D World reports that the whitepaper calls it "high partly by construction," because the 50 tasks were selected for reaching at least 80% after five demonstrations, the quantity being reported. Only 37 of the 50 succeeded in every trial, 10 had no zero-shot success, and no per-task breakdown is published. Toy-truck assembly, 65% on easier start positions, drops to 35% on the original arrangement. In mitigation, any human intervention counted as a failure.

What will bite a practitioner first is timing. The Decoder reports that during trash sorting the robot sat idle about a third of the time, waiting an average of more than seven seconds for its own planner, and that the model does not yet read touch-sensor data, which is why it misses a half-turned screw. Weights for GR00T 2 and Turbo are promised for December.

Jensen Huang's resignation

NVIDIA's 8-K of 12 November states that Jensen Huang resigned as president, chief executive and board member, effective immediately, 33 years after co-founding the company. The filing gives no reason. CFO Colette Kress is interim chief executive, and Harvey C. Jones becomes the company's first board chair. The resignation followed a week of reporting about Huang's relationship with an executive at Megaspeed, a Singapore company that buys NVIDIA chips and that the Justice Department is reportedly examining over export controls. Huang has denied compromising national security or company secrets. Senate Banking has asked him to testify on 3 December, and the Q3 earnings call is still set for 17 November.

The reason to cover this on a robotics page is how much of the field runs through one vendor. Isaac Sim and Isaac Lab are the simulators a large share of labs train in, Jetson Thor is the only compute in Agility's next Digit, GR00T is the open model many groups were waiting on, and the pending $12.93B purchase of Hugging Face, which hosts LeRobot and the GR00T integrations, is under Justice Department antitrust review. I could not find one analyst note or researcher comment on what the change means for that stack; the published reaction is all about the share price.

OpenAI folds its robotics team into general research

Fortune reported on 13 November that OpenAI has broken up its stand-alone robotics team. Robotics researchers now report to chief research officer Mark Chen, and the applied robotics and hardware groups moved into the Applied organisation under president Greg Brockman. A spokesperson confirmed the reorganisation, called robotics core to the company's long-term vision, and said general models including GPT-6 Astra now apply to robots directly. Michelle Fradin disputes Fortune's framing and says the research was merged rather than disbanded. Fortune's reporter went looking for OpenAI's booth at the robot-learning conference, number 208, and found it empty. This is the second time; the first team was wound down in 2020, with co-founder Wojciech Zaremba citing a shortage of robot data.

The technical case for the decision is Astranomy, a study NVIDIA's Seattle Robotics Lab published on 14 October. It ran GPT-6 Astra as a manipulation policy with no robot fine-tuning on RoboLab-120, a 120-task simulated pick-and-place suite with a DROID robot at 15 Hz. The strongest configuration is an agentic policy, a persistent Codex session that keeps working at the task rather than emitting one action per call.

Policy RoboLab-120 success Cost per successful episode
GPT-6 Astra, agentic Codex session 74.8% $36.40
GPT-6 Astra, direct 66.3% $1.96
Lakeside, best robot-specific model 51.0% not applicable

The agentic run burned about 30.9 million tokens per successful episode. The cheaper direct policy averaged 18.56 seconds per inference call, with the simulation paused while the model thought. Put that next to GR00T 2 Turbo's 124 ms onboard and the direct policy is about 150 times slower on the axis that decides whether a policy closes a control loop. The study shows the approach works. It does not show it running on a robot.

Embodied First Exam, from 30 authors led by Kangrui Wang, sets a bound from the other side. It runs 1,156 tasks on 47 robots both with privileged access to simulator state and under rules where the agent sees only sensor observations and cannot reset. The best agent, reported as "GPT-6 Astra in Codex" rather than as a model alone, reached 34% under the stricter rules.

Runway's Praxis-1 under Apache 2.0

Runway, known for video generation, released Praxis-1 on 11 November. The model is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and actions jointly, out with weights, inference code and both full and LoRA fine-tuning code under Apache 2.0, plus native LeRobot support. Noble Machines, Standard Bots and Ultra have run it on their own robots.

Runway reports 46.2% on RoboLab-120 across 1,200 episodes, against 42.9% for FLUX 3 Action and 28.0% for π0.5. The licence matters more than the lead, because FLUX 3 Action's terms do not permit selling what you build with it and Apache 2.0 does. The Robot Report notes that Praxis-1 ranks fifth overall on that leaderboard while Runway calls it the top-ranked world action model there, and both can be true depending on which models you count. Against GR00T 2's December weights, this is the open model a company can use today.

How much of a LIBERO score is retrieval

LIBERO is the benchmark robot papers report more than any other, a set of simulated tabletop tasks like putting a bowl on a plate. Denis Grachev, Damien LaRocque and Omar Moured audited it with the simplest method they could build, nearest-neighbour retrieval with no training at all. Take the current camera image, find the recorded demonstration frame whose embedding is closest, replay the action chunk that followed it. No gradients, no policy.

That scores 91.7 on LIBERO, where Physical Intelligence's π0.5 scores 97.7. Then the authors moved the furniture, running everything again on LIBERO-Plus, which shifts the camera, changes the lighting, swaps objects and rewords instructions.

Method LIBERO LIBERO-Plus Drop
kNN retrieval, multimodal key 91.7 82.4 9.3
π0-FAST 90.6 25.7 64.8
SmolVLA 82.2 27.9 54.3

The training-free lookup dropped least of all twelve methods tested and beat 7 of the 9 trained policies. Their explanation is the convincing part, and it is specific. Across the nine models, fine-tuning mainly makes the vision-language backbone a better retrieval key: running retrieval on a model's own fine-tuned features improved LIBERO-Plus for 7 of 8 models and beat the full policy outright for 4 of 9.

Their ask is modest and every leaderboard should grant it. Print the retrieval baseline beside the model's score, so a reader sees how much of the number is lookup.

Memory indexed by time

Some policies carry an explicit memory so they can count, which matters for a job like putting three items in a box and stopping. Chubin Zhang and Bo An show that those memories are organised by time, so the policy cannot tell waiting from working. A human teleoperator demonstrates at a steady pace, elapsed time becomes a decent proxy for progress, and the model learns that shortcut instead of the task structure.

At test time the shortcut breaks on nothing at all. A brief pause or a wrist fidget, what the authors call a null event, counts as progress, and the policy stops early. On one RoboMME counting task success fell from 47 of 50 to 16, and on another from 44 to 10. Withholding the null-event frames from memory restores performance, which localises the cause to the memory design rather than the perception or the controller. Removing the time encoding does not fix it, because null-event frames still occupy memory slots, and training with injected pauses hurts counting. Indexing memory by events recovers half or more of the loss with no retraining.

World In-Context Learning

Heng Fang and colleagues argue that in-context learning has failed on manipulation because a demonstration recorded in another scene carries no feedback about the scene in front of the robot, and their fix is to close that loop with a simulator. The frozen model starts from the single demonstration, rehearses the task in simulation, sees what went wrong, and writes what it learned into its context as a Task Memory it carries into new scenes. The implementation is called RoboScript and the code is released.

Across all 50 RoboTwin 2.0 tasks, with one demonstration each and no gradient updates, they report 83.4% mean success over clean and randomised settings, more than 20 points above the strongest trained action policy, with transfer to a physical dual-arm robot. The condition on that result is the interesting one. The frozen model only becomes good once a simulator is grading its attempts, a third answer alongside OpenAI's position that the frontier model is already the policy and NVIDIA's that the policy needs its own weights.

NHTSA's recall query on Waymo's school-bus fix

Waymo recalled its software after repeated reports of robotaxis illegally passing stopped school buses with their red lights flashing, and said the update reached every affected vehicle by 17 November 2025. On 6 November this year NHTSA's Office of Defects Investigation opened a recall query to assess whether that remedy was adequate, having found that a majority of reported incidents happened after the update. Austin Independent School District reported at least five in a letter to the agency. Footage from Atlanta shows a Waymo crossing in front of a stopped bus with its stop arm out while children were unloading, then turning left around the front of it. Waymo says it has significantly improved vehicle behaviour around school buses and has had no citations since March 2026.

A citation count is not an incident rate, and neither side has published a per-encounter number. In driving, where I work, this is the familiar hard case: rare per encounter, unacceptable every time it happens, so the only measure that means anything is normalised by exposure.

Also this week

  • Tesla opened Cybercab orders to invited customers at $29,990 plus $299 a month, deliveries from December, while NHTSA's September inquiry into how a car with no steering wheel or pedals self-certified to federal safety standards is still open. Musk says the $299 is too high.
  • SoftBank led a $500M round in Agile Robots at an $8.5B valuation, with NVIDIA, Hitachi and T.Capital, seven days after China's foreign-investment review blocked its purchase of ABB's robot business. Agile ONE has entered initial production at Fürstenfeldbruck.
  • Path Robotics raised $300M at a $2.45B valuation for Obsidian, a welding model that finds joints, plans welds and adapts mid-weld. The company says 50 welding robot dogs are on preorder for US shipyards.
  • UBTECH signed to supply 400 U1 Pro humanoids to Shanghai Yongda Holding Group, a car dealer, with the first batch targeted before the end of 2026. Treat the date as a company target.
  • The BEHAVIOR Challenge produced a disagreement with numbers on both sides. The second-place team wrote in its report that it avoided agentic harnesses and closed-source models to meet the challenge's 1 FPS floor. The ninth-place team routed 14 of the 100 tasks to a GPT-6 Astra coding agent anyway, scoring 0.639 there against 0.108 for its learned policy alone.

What we're watching

  • Does Mercedes-Benz or Apptronik publish a trial count and an intervention rate behind the 96%, and does the proof of concept become a production shift?
  • When GR00T 2's weights arrive in December, does anyone outside NVIDIA run the 50 tasks and publish the per-task breakdown?
  • Does any VLA paper or leaderboard start printing a training-free retrieval baseline beside its own score?