Physical Intelligence launches a hosted API for its robot models
A robot-model lab opens its fine-tuning stack to outsiders, and a new paper says one policy can out-assemble twenty specialists — in simulation, for now.
Physical Intelligence opened a beta of its Pi API, letting outside developers upload demonstration data, fine-tune the lab's robot foundation models on Pi's own compute, and run the result through a hosted inference gateway — no ML pipeline of their own. The company, co-founded by Sergey Levine, Chelsea Finn and Karol Hausman and valued at $5.6 billion after $1.1 billion raised, already ships its π0 and π0.5 models as open weights. The API targets everyone who looked at that repo, then looked at the data-collection and training burden behind it, and gave up. Pi says the model behind the endpoint already offers better data-efficient fine-tuning than the public weights, but names no version, no price and no latency figures — the beta is gated and onboarding in batches.
The partner list is the argument. Actor runs a wheel loader unattended on a commercial site in Utah; a dairy partner, a garment-handling shop and electronics assembly lines round it out, and most ran inference from the cloud. The sharpest detail is Almond's dual-arm Axol deburring the spacers used in its own assembly, running a model the team produced after four days of iterating with the API. That is the pitch: not a benchmark number but a turnaround time. Pi also concedes its early deployments surfaced "network connectivity challenges" — an honest admission of what a hosted robot brain means in a building with flaky Wi-Fi, and a gap competitors selling on-premise inference will attack.
A paper called AlphaNIST reports that a single reinforcement-learning policy, trained across all twenty tasks on the NIST assembly board from nothing but a binary success reward, hits 98.2% success — beating twenty separately trained specialists at 81.4%. The NVIDIA-affiliated authors (Yashraj Narang, Ankur Handa, Gavriel State among them, with Octi Zhang leading) also say dense rewards for reaching and alignment made results worse, not better: a shared policy with dense rewards managed 17.8%. The policy found its own assembly strategies — rolling an M12 nut from its edge onto a flat face, tapping an M4 nut along a bolt with one fingertip, setting an 8 mm peg down and regrasping it near the end so more of it stays exposed for insertion.
The efficiency claim deserves the emphasis the project page skips: 410 GPU-hours for the shared policy against 16,100 for twenty specialists, roughly 39 times cheaper for better aggregate success. Treat the success rate with care anyway — this is simulation, the paper is explicit that nothing here establishes robustness to perception error, latency or contact-model mismatch, and transfer to the physical board is listed as future work. Simulation-only assembly results have been impressive and unshippable before.
Industrial Next, the ex-Tesla robotics startup pitching today on TechCrunch Disrupt's Startup Battlefield stage, says its self-contained robot stations can be retooled for a new part in minutes rather than weeks, because tasks are taught in a high-fidelity simulation instead of hard-coded. Founded in 2021 by Lukas Pankau and Allen Pan — Tesla and Waymo and Momenta between them — and backed by Y Combinator and Khosla Ventures, the company announced on stage a partnership with BMW Group Technology Office to evaluate its dexterous manipulation. That partnership, as of tonight, rests on the stage announcement alone; we could find no confirmation from BMW. Pankau also says its stations already thread delicate flex cables and peel protective film off camera modules on "one of the world's most popular smartphones" — a customer he declined to name.
What to watch: whether BMW confirms that evaluation, and whether Pi publishes the model name and price behind its API.
Would you trust a robot policy that learned every assembly trick in simulation to touch your production line? Tell us in the comments.
