Builds general-purpose AI models (the π series) that let many kinds of robots see, follow language and act.
Where it stands: Partners Weave and Ultra run its π0.6 model in commercial deployments; no paying customers or revenue stated (24 Feb 2026) source
Physical Intelligence's newest general robot model, billed as an early step toward a general-purpose robot brain. It combines skills learned in different settings to attempt tasks it was never trained on, and users can coach it through unfamiliar jobs in plain language. Demos included folding laundry on a bimanual UR5e setup without task data and cooking a sweet potato in an air fryer.
A reinforcement-learning add-on for π0.6: an RL token compresses the model's internal state so a small actor-critic network can train on the robot in real time. It refines precise steps such as screwdriving, zip-tying, Ethernet insertion and plugging in power cords with as little as 15 minutes of practice, without retraining the full model.
A memory system added to π0.6, published as VLAs with Long and Short-Term Memory. A video encoder keeps the last few seconds of camera frames for short-term tracking, while short text notes record key events for the long term, so robots can stay on multi-step jobs such as cleaning a kitchen or unpacking groceries for up to fifteen minutes.
A vision-language-action model billed as a VLA that learns from experience. Using a method called Recap, it improves through its own practice attempts and teleoperator corrections rather than only copying demonstrations; demos included espresso making, folding unfamiliar laundry and assembling packaging boxes. Partners Weave and Ultra later reported running π0.6 in commercial laundry-folding and order-packing robots.