Show Notes
About twenty researchers spent a paper writing down, formally, the rule that when you tell a learned simulator to move the robot's arm, the arm should move. Six of the open-source world models the field builds on don't reliably obey, and a second paper explains why: trained only on successful demonstrations, they've never had a reason to learn what a bad action does, so feed them one and they show you the task succeeding anyway. Also: Google DeepMind's whole-body humanoid stack, which contains no world model at all, a robot that triples its success rate under a lighting change because it learned where to look, and the training examples that poison a dataset by being correct.
More from Researcher's Guide to Physical AI
Google DeepMind pitched Gemini 4 Argon as a model for every stage of building a robot; a day later Artificial Analysis put it and four other frontier models on a real arm, Argon came third, and three of the five scored zero on measuring a box that a person handles in twenty seconds. Also: RoboHarm's real-arm refusal test, where GPT-6 Astra carried out 60 of 100 dangerous instructions, NVIDIA's AlphaNIST learning 20 assembly tasks from a yes-or-no reward, Orbia catching world models moving the lamp, Physical Intelligence renting its model, and Agility's vote date.
Five frontier AI models were put in charge of robots across 84 tasks, and 63 went unsolved by all of them; one announced its job was done and then kept raising and lowering an empty arm. Marc Raibert and Agility's Jonathan Hurst say today's success rates aren't close to good enough. Also: Meta's FAIR scales RoboJEPA to 8B with a checkable scaling law, DreamTrue trains a world model on failures it invents, Waymo borrows $5B against a million paid rides a week, and Boston Dynamics hires Amazon's Rohit Prasad as CEO.
Boston Dynamics taped its engineers' pinkies to their ring fingers for a day, then shipped Atlas a four-fingered hand built for manufacturing, repair and simulation, and a BTIG count says only 13% of this year's humanoids walk on two legs. Tsinghua's Leap Lab rebuilt both kinds of world-action model on one budget: the fast latent kind scored 6% where the explicit kind scored 70% on tasks learned from video, and the fix, Simple-WAM, keeps one step of noise and drops the rendering. Also: Nucleus says a human drove 40% of its factory demo, and California will make employers name the machine that replaced their workers.
Black Forest Labs, the company behind the FLUX image models, put out the best open robot policy going and then paid an outside firm to run it blind on an arm it doesn't own: 28 of 30, against 13 for pi-0.5. Figure's home-robot success rate turned out to have a denominator all along, 237 of 420, and a rival CEO read it straight back at them. Also: a group in Beijing traced the physics failures in video world models to a positional encoding borrowed from language, fixable by rescaling one frequency.
Dario Amodei asked the frontier labs to pace themselves, Sam Altman said OpenAI would match one commitment, and two days later Jensen Huang told the President on speakerphone "we're not going to let that happen, sir." The essay never mentions robots, though nearly everyone on both sides of it owns a stake in one, and nobody said whether a robot foundation model is part of what gets paced. Then Figure sent its humanoids into thirty homes they had never seen, where pretraining on people's phone videos did most of the work, with no trial count published. Also: human touch standing in for robot touch, Digit 5's new knees, and UBTECH's ten-minute humanoid factory.