Frontier models fail humanoid control benchmark

Today: a held out benchmark frontier models fail, a 500 hour human dataset, Hitachi and FANUC on factory floor learning, ABB’s new compact robots, a 50 million dollar data offer, and a robotaxi petition review.

Robo-bench finds the arm partly reachable and the humanoid out of reach

Vals AI published Robo-bench, a held out evaluation of general purpose agents controlling simulated robots: 24 tasks, 13 on a 7 degree of freedom Franka Emika Panda arm and 11 on a 29 degree of freedom Unitree G1 humanoid with two Dex3 hands.

Both run in MuJoCo held to their real limits: joint ranges stop where the hardware stops, arm targets are rate limited to datasheet speeds, and humanoid motors follow a torque speed curve. The humanoid ships with no balance or walking controller, so the agent commands every motor including the legs, and it gets only the datasheet, sensors and cameras, never the simulator. Episodes last three minutes of wall clock time with physics live throughout, so thinking spends the episode.

The results are blunt. GPT-6.1 Sol leads at 28.33 percent accuracy, averaging 51 out of 100 on the arm tasks and solving four of 13, while no model solves a single humanoid task and none averages more than 2 out of 100 on them, and two models, Grok 4.7 and Gemini 4 Argon, sent no motor command at all in 16 and 15 of their 24 episodes. Sol took 79.89 seconds to its first move; in the balance scale task it wrote its own controller and sent 2,204 joint commands.

Balance needs a correction roughly once per second, faster than a model turn, so an agent that succeeds has to emit code that closes the loop itself: the code as policy pattern with a stopwatch attached. Note the contrast with vendor benchmarks. Google claimed 87 percent for Gemini 4 Argon on embodied reasoning, reported yesterday, while the same model scored 0.83 percent here, on a test of closed loop control under time pressure.

HumanVerse turns 500 hours of human motion into a humanoid policy

A tech report from the Lambda-humanoid team, arXiv:2610.18950, describes HumanVerse-500: 500 hours of open world human activity captured with synchronized egocentric video, whole body motion and hand motion, using a neck mounted GoPro matched to the robot camera height, wearable PICO tracking for 24 body joints, and image based reconstruction for 21 joints per hand (project page).

The pipeline is the contribution. It aligns the human camera view and arm appearance to the robot, corrects the human motion reference, fits a robot reference through whole body inverse kinematics, then refines trajectories with dynamics aware optimization. The model, lambda zero, transfers zero shot to a Unitree G1 without target task robot demonstrations: 76.2 percent progress on long range loco-manipulation and 51.7 percent across four whole body tasks, with the paper reporting 90.0 percent success on SIMPLE and 62.5 percent on real world loco-manipulation.

Collection economics get a number too: about 70 seconds of robot time against 33 of human time per demo for whole body manipulation, and 130 against 25 seconds for long range work, roughly 3.4 times throughput. Whole body human data, not just hands, is what the mid-training stage consumes, which makes this a direct answer to the robot data problem.

Hitachi and FANUC put physical AI on the factory floor

Hitachi and FANUC announced a strategic partnership on 30 September to commercialize physical AI, combining Hitachi’s HMAX Industry models with FANUC’s AI equipped industrial robots (FANUC, Hitachi).

The mechanism matters: Hitachi Group factories act as “Customer Zero”, starting at its facilities in the Ibaraki region, so the AI learns site specific production rules instead of relying only on imitation learning or digital twins. The first applications are picking parts of varying shapes and changeovers when production items change, scored on recognition accuracy, robot motion, takt time and quality. The partners also plan to test Hitachi’s edge AI semiconductor with FANUC robots, with joint customer deployment from fiscal 2027.

For developers this is a counterexample to benchmark chasing: the claim is not a higher score on a fixed task list, it is a tighter loop between production data, model updates and deployed cells, which is what buyers compare against changeover downtime.

ABB ships the IRB 1700 family for precision work

ABB Robotics introduced the IRB 1700 and IRB 1700ID on 12 October, showing them at the China International Industry Fair in Shanghai. The company claims up to 20 percent shorter cycle times than comparable robots, path repeatability down to 0.03 mm, joint speeds up to 700 degrees per second, 15 percent lower power consumption than the previous IRB 1600, service intervals of 20,000 hours between oil changes, and a footprint 50 percent smaller than its predecessor (ABB, ASSEMBLY Magazine). Treat those as vendor figures.

Five variants span 6 to 12 kg payload, the product page lists the IRB 1700-12/1.45 and IRB 1700-10/1.65, and both run on OmniCore controllers, which also carry RobotStudio, AI Robot Trainer, EyeMotion and automatic collision free path planning, so classical motion and learned perception share one deployment path. Listed applications include material handling, arc welding, deburring and machine tending.

Humxn offers 50 million dollars for real work data

HUMXN Research Labs launched an initiative to acquire and license 50 million dollars of data from skilled trades, farming and manufacturing businesses (Newsfile release). It is buying service records, repair histories, job site photos and training videos, and says it pays about 20 dollars per hour of approved wearable camera footage. The company reviews submissions, declines incomplete or duplicated data, verifies ownership and de-identifies personal information.

That is its own statement, with no independent confirmation of deal sizes. The reason to watch it is the price signal: if an hour of real work footage settles near 20 dollars, that is the number synthetic pipelines have to beat, and the simulation route is cheaper per frame but weaker on contact fidelity.

NHTSA reviews a robotaxi school zone petition

NHTSA began reviewing a defect petition on 9 October after a late September filing alleging that roughly 1,000 Tesla Model Y Robotaxis and Cybercabs exceed posted speed limits in active school zones (ExtremeTech). The filing reportedly contains three video links tied to four incidents: one complaint filed with the agency and three social media reports, with no crashes, injuries or deaths mentioned.

Nothing is established yet. Anyone can submit a petition, the agency has 120 days to review the allegations and decide whether to open a formal investigation, and a review is not a finding of defect. Tesla already drew an examination of Full Self Driving behaviour after traffic violation reports in 2024, and regulators are separately reviewing technical data behind the Cybercab, which ships without a steering wheel or pedals.

The engineering point: compliance with time and location dependent rules is a different capability from collision avoidance, and it is increasingly the test written down first. It is the gap a gate in front of the model exists to close.

Quick Bites

• Rivian backed Mind Robotics hired Doug Field, formerly of Ford, Tesla and Apple’s car project, as president. The venture was created by Rivian CEO RJ Scaringe, has more than 1 billion dollars in backing, and uses Rivian’s factory as its first customer (TechCrunch).
• NYSOFA extended its Intuition Robotics partnership for three years with 2.1 million dollars for the ElliQ companion robot. Reported program data for 2026: 92 percent find it useful, average engagement of 48 interactions per day against 29 in earlier reporting, and 94 percent reporting less loneliness (The Robot Report).
• Open Robotics HQ says six local ROSCons land in the final quarter of 2026, alongside work on OSRA membership paths, an incoming technical governance committee meeting and the ongoing yearly audit (ROS Discourse).