Physical Intelligence π-0

À propos de Physical Intelligence (PI)

Fondée 2024 · San Francisco, California, USA

Physical Intelligence (PI) is a San Francisco robot-foundation-model startup founded in 2024 by Sergey Levine, Chelsea Finn, Karol Hausman, Brian Ichter, Suraj Nair, Lachy Groom and others, with the explicit mission of building general-purpose foundation models for robots. The company raised $400M in early 2025 (led by Jeff Bezos, Thrive Capital, Lux Capital and others) at a multi-billion-dollar valuation. π-0 ('pi-zero') is PI's flagship VLA model, announced in October 2024 as a generalist robot policy that can drive many different robots through dexterous manipulation tasks - folding laundry, bussing tables, packing boxes - using a single set of weights and natural-language instructions. The model is published with a public technical report, and a subset of weights and code are released under the open-source 'openpi' GitHub repository to support reproducible research.

Architecture

Flow-matching Vision-Language-Action policy

π-0 is a Vision-Language-Action model built on a 3B-parameter PaliGemma multimodal backbone (Gemma LLM + SigLIP vision tower), with a dedicated 'action expert' transformer head trained as a flow-matching policy. Inputs are multi-view RGB images, robot proprioception, and a natural-language instruction. The action expert outputs continuous action chunks at high frequency via flow matching - a recently popular alternative to diffusion that produces smoother trajectories with fewer integration steps. π-0 is pretrained on a large dataset of ~10,000 hours of multi-embodiment robot teleoperation collected by PI and partners, plus Open-X-Embodiment data and offline simulation. Post-training fine-tuning specialises the policy for individual robots and tasks. The model handles cross-embodiment action spaces (Aloha bimanual, Franka, mobile manipulators, humanoids) through learned action-space adapters.

Paramètres

~3B (PaliGemma backbone + action expert)

Capacités

• Generalist VLA driving multiple robot embodiments
• Flow-matching action head producing smooth continuous actions
• 3B PaliGemma backbone (Gemma LLM + SigLIP vision)
• Natural-language instruction conditioning
• Long-horizon dexterous manipulation (laundry folding, bussing tables)
• Multi-view + proprioception inputs
• Reactive closed-loop control at high frequency
• Partially open-sourced via the openpi GitHub repo
• Best for: dexterous manipulation research, multi-task generalist policies.

Entraînement et licence

Approximately 10,000 hours of curated teleoperation data on a fleet of Physical Intelligence robots, augmented with Open-X-Embodiment cross-embodiment data and offline simulation. Multi-view RGB + proprioception + natural-language instructions.

Licence: Partially open-source - core π-0 model weights and inference code released via the openpi GitHub repository under permissive terms for research; commercial deployment is governed by Physical Intelligence's own licence and partnerships.

Tests de sécurité: PI emphasises safe behaviour shaping during data collection and uses force-limited, workspace-constrained robots in deployment. No formal RSP-style policy is published; safety guarantees rely on hardware constraints, supervised testing and downstream controllers.