Physical Intelligence

What is Physical Intelligence?

Physical Intelligence develops learning algorithms for a general model that can control any robot to do any task. Its work is published as a series of vision-language-action models, from the first generalist policy, π0, through π0.5 and π*0.6 to π0.7.

Key capabilities of Physical Intelligence

• Generalist robot policies: one model family aimed at many robots and tasks.
• Open-world generalization: π0.5 targets unfamiliar settings such as cleaning kitchens and bedrooms it has not seen.
• Learning from experience: π*0.6 uses reinforcement learning to improve from experience.
• Steerable behavior: π0.7 is described as a steerable model with emergent capabilities.
• Open weights: weights and code for π0 and π0-FAST were released in February 2025.
• Pi API: upload data, fine-tune models and run cloud-hosted inference on your own robots.

How Physical Intelligence works

Physical Intelligence trains models that take camera views and language instructions and output robot actions. Developers can use the open π0 and π0-FAST releases, or request access to the Pi API, where they upload data, fine-tune and run hosted inference. The API is in beta and access is by waitlist.

Who uses Physical Intelligence?

Robot builders and developers who want a pretrained policy rather than training control from scratch. The company's site names no other audience.

Physical Intelligence pricing

Physical Intelligence does not publish pricing or license terms. The Pi API is in beta with a waitlist, so cost is arranged with the vendor.

Physical Intelligence alternatives

Skild AI also builds a general-purpose robot brain, so the two are direct peers in robot foundation models. fal is a hosted inference platform for generative media models, not a robotics policy provider. Modular MAX is an inference and serving stack for AI models, used to run models rather than to supply robot policies.