Why GPT-6 Astra is suddenly so good at robotics

OpenAI's GPT-6 Astra, released on 3 September as a frontier model for computer use and coding, outperformed all 40 trained robot policies on RoboDojo's benchmark of 42 simulated tasks with no robot-specific fine-tuning, averaging a 22.48% success rate versus 0.88% for GPT-5.5 using the same single gripper action. A follow-up harness, PyRUA-Lean, prompted Astra to generate code rather than control the gripper directly and averaged 51.6%, beating the leaderboard's top entry at the time by 19 points. The results arrive as Physical Intelligence raises a $20 billion round, and Altimeter's Freda Duan called Astra robotics' "GPT-3 moment.

GPT-6 Astra drives robots now, which is awkward timing for Physical Intelligence's $20 billion round. In October, researchers at the University of Illinois pointed a camera at a transparent drawer and asked a coding agent to write code that would get a robot arm to open it. Success: 4 out of 10. Then they gave it the force sensor data from the wrist, and it iterated on its Python program https://arxiv.org/html/2610.31841v1 , revising it after every failure. Success: 10 out of 10. Then they took the coding agent away, and the program it had written kept opening the drawer on its own. Roboticists have a name for why this shouldn't work: Moravec's paradox https://en.wikipedia.org/wiki/Moravec%27s paradox , named after Hans Moravec, who in 1988 observed that "it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility". In robotics it has looked prophetic. In 2022 SayCan's planner could plan an action sequence for all 21 user requests https://arxiv.org/pdf/2204.01691 involving drawers, but its visuomotor policy could only execute about a third of them, despite being trained on 68,000 demonstrations collected from humans For the last three years, generalist robot manipulation has been an arms race among the largest labs to collect as many human demonstrations as possible to train ever more capable visuomotor policies, the hands, with comparatively less attention paid to the planners, the brains. Vision-language-action models VLAs learn to turn camera images and an instruction straight into motor commands, and Figure's Helix https://www.figure.ai/news/helix even borrows Kahneman's names for the two halves: a slow System 2 that decides what to do and a fast System 1 that moves the arms. In Illinois they took an off-the-shelf System 2 Claude Fable 5, from Anthropic and had it write System 1 on its own. This strikes me as the right build order, and the labs as having had it backwards. Learned policies are great, but mostly where it's hard to find a good metric for success. Folding laundry is one of those. Opening a drawer is not. Everything is a drawer except for laundry. Astra drives robots now RoboDojo added GPT-6 Astra https://openai.com/index/gpt-6-astra/ , released by OpenAI on 3 September as a frontier model for computer use and coding, to their benchmark https://arxiv.org/html/2609.24170v1 of 42 simulated tasks. With no robot-specific fine-tuning and only one action, move the gripper, it outperformed all 40 trained robot policies on the leaderboard, with an average success rate of 22.48%. GPT-5.5, with the same single action, averaged 0.88%. It was aided by recipes task descriptions without action-level guidance , and the simulator paused while it deliberated in one study https://arxiv.org/html/2610.22471 of Astra on RoboDojo, an average of 55 minutes of computation for at most 76 seconds of motion . Within days, trained models surpassed it. PyRUA-Lean next prompted Astra to generate code rather than directly control the gripper. Despite lacking both recipes and a learned policy, this harness averaged 51.6% https://arxiv.org/html/2610.01939 , beating the top of the leaderboard at the time by 19 points. By changing the harness around the same model, and raising its reasoning effort, performance more than doubled. Here's the rule I'd tattoo on every robot founder: if the communication between the brain and the hands is in words or code, then it's easy to swap out the brain without having to retrain the hands; if the communication is in embeddings numbers that only have meaning to the neural network you've trained on them , then every time you get a better brain you're probably going to have to retrain the hands. Altimeter's Freda Duan calls Astra robotics' GPT-3 moment https://x.com/FredaDuan/status/2116727410972590380 , and she says some in the robotics community seem "reluctant to acknowledge it". I'm not one of them: the most important robot brain of 2026 was built by a chatbot company. Jim Fan's lab switched to Astra the day it came out: Self-reported, admittedly. In July, I wrote https://josef.cn/blog/turks-all-the-way-down about a control reflex for robots: notice when the policy is about to fail, and ask a larger AI model for help. I had the build order backwards In October the big model went first and wrote the robot's program itself; in RAX the reflex is still there, to catch exceptional cases and call the model back in. The chef writes the recipe and leaves The VLA labs are trying to replace the robot's program. These agents are after the job of the robot programmer, that specialist who's been hired by manufacturers for decades to program a robot's behaviours, sign it off and leave it be, on a robot that lasts 12 to 15 years https://ifr.org/img/worldrobotics/Sources Methods WR 2025 Industrial Robots.pdf . Jeff Mahler, Ken Goldberg's co-founder at Ambi Robotics, has already used an agentic harness https://x.com/Ken Goldberg/status/2100986412762087909 on a real industrial problem and deployed the result to robots that sort parcels all over America. RAX https://arxiv.org/html/2610.22496 demonstrates the same loop on a Franka arm and an air fryer. After 72 minutes of inference, the agent had figured out how to open, load and close the air fryer. It then wrote some Python about 1,500 lines that does the same thing with only 18 seconds of compute in a three-minute run It ran the program nine times, revising it as it went, and by its own measurements four of the nine runs were clean. The whole session ran from 09:18 to 12:05 one morning. Our first robot was held together with zip ties and delusion, and it took three engineers to get it to version 547 still serving thousands of meals a month as of last year . RAX's agent debugged its own code in a morning. On the air-fryer program's first run, though, an awkward arm posture banged the robot's elbow into a camera: RAX includes several software checks https://arxiv.org/html/2610.22496 , but none of them protects the cameras. And models seem fine with attempting actions they shouldn't be allowed to try: Where the hands still win I've been peeling potatoes since I was six, in my parents' restaurant. I don't know how much force I use with the peeler, and I doubt I keep it in any particular range; if I had to guess, I'd say my policy is pretty learned. Illinois's agent peeled a cucumber held by a person https://arxiv.org/html/2610.31841v1 with code that reads a force sensor on the robot's wrist 100 times a second, tries to keep the force in the 4.5 to 7.5 N range, and stops the blade when it loses contact with the cucumber which happened 2.2% of the time, compared to 40.6% for the camera-only agent . I could have coded that up myself That force sensor is like having a test suite for a coding agent, as it lets the agent know when its program screwed up. That's the line between the drawer and the laundry, and it's the question my own research keeps coming back to. Cloth manipulation has so far been the domain of learned policies, and Physical Intelligence's laundry folding https://www.pi.website/blog/pistar06 is the strongest case against everything I'm arguing. Awkward timing for $20 billion Where the brain stays in the loop, it borrows the hands. On EBench https://arxiv.org/html/2610.35832v1 , Astra could hand steps to Physical Intelligence's π0.5, and its success rate went from 46.7% to 53.0%, better than either component alone, though still short of the best trained policy on EBench's full leaderboard 61.4% . Swap π0.5 for a rival policy, InternW0-Δ, and the same brain scored 56.0%, so the hands are swappable too. Now the round. There's no revenue at PI https://qz.com/physical-intelligence-funding-talks-20-billion-valuation-102726 , but it's reportedly in talks https://bloomberg.com/news/articles/2026-10-27/robotics-lab-physical-intelligence-in-talks-for-a-20-billion-valuation at a valuation of about $20 billion. That's a price for a company that owns the robot's brain, and after this autumn I don't think PI does. Stone Tao, a researcher at PI, wrote in September https://x.com/Stone Tao/status/2099398636870283360 about "the feeling that Astra could just zero shot all the robotics foundation model companies", and then that one could "leverage astra together with a good robotics model to get even better results". This second part is EBench, and it's PI's best case scenario: being the best set of hands a rented brain can buy. But valuing a company at $20 billion assumes that laundry-like tasks dominate the jobs robots do for pay. The jobs robots are hired to do in this essay sorting parcels, assembling electronics seem more drawer-like to me. Why the labs buy so much data A learned policy is only as good as its training data, and VLAs need data at every step of the way. For pretraining, labs lean on data that doesn't come from their own robots: Physical Intelligence's π0 https://arxiv.org/html/2410.24164 began with PaliGemma, a 3-billion-parameter VLM pretrained on internet data its own pretraining then added over 10,000 hours of robot data , and Dyna pretrained its DYNA-2 model https://prnewswire.com/news-releases/dyna-robotics-unveils-dyna-2-world-action-model-demonstrating-first-true-scaling-law-in-robotics-powered-entirely-by-human-data-302847114.html on "1 million hours of human video". Post-training requires actual robot data, much of it from teleoperation. Business Insider describes armies of human operators https://businessinsider.com/1x-humanoid-robot-training-humans-world-models-optimus-rival-2026-1 and data collectors who perform tasks from squatting to washing dishes while being recorded or wearing sensors, or teleoperate robots in VR headsets and motion-capture suits, for pay that starts at $25 an hour at Tesla and 1X. Encord, which sells robot training data, told Business Insider https://businessinsider.com/ai-companies-human-workers-robot-training-data-2026-10 in October that demand is roughly 10 times higher than a year ago, with customers requesting 10,000 or even 100,000 hours of data instead of the 1,000 hours they asked for before. The research is moving away from the operator. 1X's CEO says the big unlock https://businessinsider.com/1x-humanoid-robot-training-humans-world-models-optimus-rival-2026-1 is that intelligence now "scales with the number of deployed robots, instead of the number of operators you have gathering data", and RAX's Yilun Du asks https://x.com/du yilun/status/2111143189222404427 "What happens when we scale EXPLORATION for robot agents?" Agentic robotics makes the same bet, that robot intelligence scales with practice. A frontier model arrives already pretrained by someone else , forgoes robot-specific post-training, and instead learns from its own practice on the real robot, storing that experience as code. Remember SayCan's 68,000 demonstrations? Four of its authors went on to co-found Physical Intelligence https://pulse2.com/physical-intelligence-raises-400-million-at-2-4-billion-valuation-to-develop-ai-for-robotics/ . FORCA requires "no demonstrations, task-specific training, or simulation". Some unsolicited advice for roboticists: stop collecting a thousand more demos and start by putting a force sensor on the wrist. If there's a sensor that can tell you whether you succeeded, you're competing with agents that can write air-fryer code in a morning and mostly just cost tokens. Some unsolicited advice for VCs looking at robot-brain startups: understand the fraction of the work customers pay for that sensors can validate, and the fraction that actually requires a learned model or a human to judge. Let agents have the drawers, and value the startup on the laundry. Rodney Brooks recently called the rush to collect robot data sets "a lot of wasted capital" https://www.automate.org/automated-podcast/episodes/rodney-brooks-2 . I agree, and I think Physical Intelligence and the VLA startups priced like it are overblown: why teach robot hands what a sensor can already check? The big opportunity is the robot programmer's job, which a rented System 2 can now do. Moravec's paradox still holds for laundry. For tasks a sensor can check, it cracked this autumn. Note: OpenAI hasn't said whether Astra was trained on any robot data. The RoboDojo paper says there was "no robot-specific or task-specific fine-tuning performed for this evaluation", but whether some of RoboDojo's material made it into Astra's training data is, the authors write, "unquantified rather than ruled out". Braindump by Josef Chen Originally published at josef.cn https://josef.cn/blog/agentic-robotics . Cover: Johannes Vermeer, The Lacemaker, c. 1669 to 1671, Musée du Louvre.