Advertise on podcast: 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. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/011/ More episodes and the complete archive: https://physicalaipodcast.co

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. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/010/ More episodes and the complete archive: https://physicalaipodcast.co

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. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/009/ More episodes and the complete archive: https://physicalaipodcast.co

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. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/008/ More episodes and the complete archive: https://physicalaipodcast.co

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. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/007/ More episodes and the complete archive: https://physicalaipodcast.co

A humanoid maker's books went public for the first time under securities law, and they show $1.8 million of robot sales against a $2.5 billion merger valuation, while the company selling robot software rather than robots reported fifty-five times that in recurring revenue. In the same seven days, three groups published three incompatible answers to where a robot's imagination should come from, and none of them cites the others. Also: a model trained to edit pictures turns out to be better at depth than the models built for depth, and four tactile datasets landed at once, with the largest still short of what its own authors say the field needs. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/006/ More episodes and the complete archive: https://physicalaipodcast.co

NVIDIA agreed to buy Hugging Face for just under thirteen billion dollars, which puts the shelf almost every open robot model sits on inside the company that sells the chips those models train on. Its own announcement of the deal never uses the word robotics, not once, and three days earlier a small company in Oslo opened a graded, licensed marketplace for humanoid training data in Hugging Face's robot file format. Also: Skild claims one video in the context window can replace the fine-tuning step entirely, a policy graded on the force it expects to feel goes from fifteen percent to eighty-two on sub-millimetre assembly, and Waymo opened three cities on the minivan it built to be cheap in the same week Tesla started charging riders in forty-five Cybercabs. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/005/ More episodes and the complete archive: https://physicalaipodcast.co

Anthropic previewed a driver specification that would let any model discover and operate a microscope, a laser or a robot arm, and in the limitations section of the same announcement it says the model can't reliably predict that shaking a liquid makes bubbles. On the same day, a federal judge threw out the Pentagon's blacklisting of the company for refusing to let its models drive autonomous weapons. Also: Unitree gave back half its value while a rival raised close to a billion dollars in the same five days, the Beijing robot games made the flat sprints autonomous and quietly left the hurdles teleoperable, and a sprinting humanoid invented a new arm posture because swinging its arms would have cooked its shoulders. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/004/ More episodes and the complete archive: https://physicalaipodcast.co

Anthropic has told its bankers it expects the largest first-time share sale in history, and it is reportedly still trying to buy a world-model company for six billion dollars while it does. A private lab can fund a five-year bet on physics, and a public one answers to a quarterly clock, which makes this the most consequential open question in Physical AI right now. Also: Unitree closed its first day of trading up 460 percent on a shipment number three organisations can't agree on within fifty percent, three labs independently decide the headroom is in the loop around the policy rather than in the weights, and a new benchmark finds that up to a quarter of recorded successes on soft objects crushed the object. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/003/ More episodes and the complete archive: https://physicalaipodcast.co

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. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/002/ More episodes and the complete archive: https://physicalaipodcast.co

Two teams tested, independently and in the same week, whether the video-generating world models robotics is betting on actually understand physics. They get the shape of a falling mug right and the numbers underneath it wrong, which matters, because these are the simulators meant to replace real robot data. Also: why choosing your training video beats collecting more of it, a robot that spots a fake instruction and obeys it anyway, and a short primer on world action models, including why imagining the future turns out to matter in training but not at run time. Full write-up for this episode, every number, method detail and citation, at https://physicalaipodcast.co/episodes/001/ More episodes and the complete archive: https://physicalaipodcast.co