π0.7 shows emergent robot skills
Physical Intelligence unveiled π0.7, a robotics foundation model that the company says demonstrates emergent skills like zero‑shot shirt folding and operating household appliances via language prompts. The posts highlight cross‑embodiment training as the method behind those zero‑shot capabilities. (x.com) (x.com)
Robots usually learn like apprentices on one machine at a time; Physical Intelligence said on April 16 that its new π0.7 model can transfer skills across robots and follow new language instructions on tasks it was not directly taught. (pi.website) Physical Intelligence described π0.7 as a “general-purpose” model and said it matched fine-tuned specialist systems on dexterous jobs while also handling unseen tasks such as new kitchen appliances and laundry folding on a different robot. The company published the announcement and paper on April 16, 2026. (pi.website) The company said the model’s shirt-folding demo used a robot that had no laundry-folding training data, and its appliance demos included operating unfamiliar household machines through spoken or typed prompts. TechCrunch reported the air-fryer example drew on only two loosely related episodes in the training data, according to the company’s researchers. (pi.website) (techcrunch.com) A robotics foundation model works like a large language model for motion: it takes in camera views and instructions, then outputs motor commands. Physical Intelligence’s earlier π0 system, published on October 31, 2024, was built to control multiple robots from images, text, and action data instead of hand-coded scripts. (pi.website) The main bottleneck is data. Physical Intelligence said in 2024 that there is no internet-scale archive of robot experience, so its answer was to train one policy on mixed data from many robots rather than collect a fresh dataset for every arm, gripper, and task. (pi.website) That method is called cross-embodiment training: combine demonstrations from different robot bodies into one model so a skill learned on one machine can help another. In its 2024 π0 paper, the company said it used data from 7 distinct robot configurations and 68 tasks; its blog described the broader training mix as spanning 8 distinct robots. (arxiv.org) (pi.website) π0.7 adds what the company calls “diverse context” in the prompt. The prompt can include plain-language instructions, subgoal images, and episode metadata that tell the model not just the task, but the strategy or quality target for doing it. (pi.website 1) (pi.website 2) Physical Intelligence said that extra context lets π0.7 use a wider mix of sources, including robot demonstrations, autonomous runs that include failures, human data, and non-robot data. The company said that richer prompting is what enabled the zero-shot and cross-robot behaviors in its tests. (pi.website 1) (pi.website 2) The claim lands as investors keep pouring money into general-purpose robotics software. Bloomberg reported in November 2025 that Physical Intelligence raised $600 million at a $5.6 billion valuation, and the company’s open-source repository says its earlier models were released with checkpoints trained on more than 10,000 hours of robot data. (bloomberg.com) (github.com) The next test is outside the company’s own demos: whether other labs can reproduce the same transfer on their hardware and failure cases. For now, π0.7 is a research result pointing at a robot model that needs less task-by-task retraining than the systems that came before it. (github.com) (pi.website)
