Physical Intelligence Ships π0.7, Robot Brain That Folds Shirts with Zero Training Data

Physical Intelligence, the Silicon Valley robotics foundation-model startup backed by Stripe veteran Lachy Groom, shipped π0.7 on April 16, 2026. The new architecture moves past simple imitation into compositional generalization: the ability to “mix and match” learned concepts to solve problems it was never trained on.

The headline demo

Physical Intelligence put π0.7 on a UR5e bimanual robot it had never seen, pointed it at a pile of laundry, and asked it to fold shirts. Zero training data for the shirt-folding task on that specific hardware.

The result: π0.7 achieved a success rate that matched expert human teleoperators attempting the task for the first time on the same hardware.

That’s the kind of generalization gap that separates “a robot you have to train for each task” from “a robot brain you point at a new task and it figures it out.” Robotics research has been chasing this for decades.

What changed architecturally

π0.7 combines three things no prior generalist robotics model did together:

1. Emergent compositional generalization. Given language instructions for a novel task, the model breaks it into sub-skills it already knows and composes them on the fly.

2. Multimodal prompting framework. Language coaching (natural language task guidance), metadata, and visual subgoals generated by a lightweight world model work together to steer robot behavior.

3. Steerability at inference time. Operators can adjust behavior mid-task without retraining. The model responds to new language commands and visual feedback as the task unfolds.

Benchmark results

π0.7 as a single general-purpose model matched or exceeded the RL-tuned π*0.6 specialist models across:

• Espresso making (multi-step barista task)
• Box assembly (dexterous manipulation with varying part orientations)
• Shirt folding on UR5e (zero prior training)

Previously these each required their own specialist model or extensive fine-tuning. π0.7 is one model for all of them, with the added benefit of generalizing to new tasks.

Why this matters

For robotics: Compositional generalization is what separates industrial automation (program each task explicitly) from general-purpose robotics (point the robot at a task and it works). π0.7 is the clearest published step toward the latter.

For AI broadly: The same compositional-generalization challenge exists in LLM agents working across new toolchains, visual agents working on new interfaces, and any system trying to handle tasks outside its training distribution. Techniques from π0.7 will likely cross-pollinate.

For the humanoid race: Physical Intelligence is building foundation models, not humanoid bodies. Their model can run on a UR5e, a humanoid torso from Figure or Apptronik, or any compatible robot. The foundation-model layer is the strategic position that lets multiple hardware partners plug in.

The Silicon Valley context

Physical Intelligence is one of the most-discussed robotics startups in Silicon Valley as of April 2026. Comparable positioning to Physical Intelligence’s:

• Figure AI (humanoid robots, Nvidia / OpenAI-backed)
• 1X Technologies (NEO humanoid, OpenAI-backed)
• Skild AI (general-purpose robot brain)
• Covariant (warehouse / industrial generalist)

The π0 → π0.6 → π0.7 release cadence has been steady, with each model showing meaningful capability step-ups.

Open questions

• Safety. A robot that generalizes means a robot that can also misunderstand novel situations in novel ways. Failure modes on tasks it was never trained for are hard to characterize.
• Robustness. Folding one pile of shirts in a lab is different from folding 1,000 real-world varied laundry situations over time. Field deployment results pending.
• Access. Physical Intelligence has not publicly opened π0.7 to third-party developers. Early partners have access; broader release timeline unclear.
• Scaling. The emergent compositional generalization showed up at current model scale. Whether further scaling continues the trend or plateaus is the next research frontier.

Sources

• π0.7: a Steerable Model with Emergent Capabilities - Physical Intelligence
• TechCrunch: Physical Intelligence says its new robot brain can figure out tasks it was never taught
• Humanoids Daily: Physical Intelligence Unveils π0.7
• Digital Today: Physical Intelligence unveils pi 0.7