π0.7 Robot Brain Unveiled
- Physical Intelligence unveiled π0.7, a robot brain that combines language, vision, and action models for untrained tasks. - The team reports roughly 80% success folding laundry and demonstrated compositional generalization across different robots. - The model uses a 4B‑parameter language backbone and claims transferability, indicating movement toward generalist robotic stacks (x.com).
Robots usually learn the way a factory worker learns one station: one task, one setup, one script. Physical Intelligence says its new π0.7 model can take plain-language instructions and carry out unfamiliar physical tasks without task-specific retraining. (pi.website) Physical Intelligence published π0.7 on April 16, 2026, alongside a paper describing it as a “steerable” robot foundation model trained to work across tasks, environments, and robot bodies. The company says the same model can handle dexterous jobs, follow new commands, and transfer skills between different machines. (pi.website) The system is a vision-language-action model, which means it takes in camera views and text prompts, then outputs motor commands for a robot. In the paper, the team says π0.7 is prompted not just with a task description but also with subgoal images and episode metadata that describe how the task should be done. (arxiv.org) That extra context is the core claim. Physical Intelligence says richer prompts let π0.7 mix robot demonstrations, autonomous runs, and non-robot data instead of relying only on examples collected for one robot doing one job. (pi.website) The headline demo is laundry. The company says it collected laundry-folding data on one static bimanual robot, then evaluated π0.7 on a different bimanual UR5e system with no laundry-folding training data from that hardware. (pi.website) The paper and blog frame that result as “zero-shot cross-embodiment generalization,” robotics jargon for one robot using skills learned from another robot with a different body. The same release also highlights unseen kitchen-appliance tasks, including operating an espresso machine and using an air fryer after language coaching. (arxiv.org) (techcrunch.com) Physical Intelligence is not starting from scratch here. It introduced π0 in October 2024 as a generalist policy trained on data from seven robot configurations and 68 tasks, then released π0.5 for broader open-world generalization and π*0.6 for reinforcement-learning fine-tuning on hard manipulation jobs like espresso making and laundry folding. (pi.website 1) (pi.website 2) (pi.website 3) π0.7 is the company’s argument that one model can now do more of that work out of the box. In its April 16 post, Physical Intelligence says π0.7 matched the speed and robustness of its fine-tuned specialist π* models on some dexterous tasks while generalizing better across scenes, tasks, and robot platforms. (pi.website) The broader problem is data. Physical Intelligence said in its 2024 π0 release that robotics lacks the giant shared datasets that helped large language models, so new robot skills usually require collecting fresh demonstrations on the exact machine and task being targeted. (pi.website) That is why transfer claims get attention in robotics labs and startups. If one policy can reuse what it learned about drawers, cloth, grippers, and appliances across many robots, companies can spend less time recording demonstrations for every new deployment. (pi.website 1) (pi.website 2) The open question is how far the generalization goes outside company-selected demos. Physical Intelligence’s paper reports the experiments, but independent replication will decide whether π0.7 is a lab result, a deployable stack, or the start of a broader shift toward general-purpose robot control. (arxiv.org) (techcrunch.com)
