Physical Intelligence π-0.5

Par Physical Intelligence (PI)

Dibināts 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 and co-founders. π-0.5 is the upgraded successor to π-0, announced in April 2025, focused on open-world generalisation: the same robot policy is shown deploying in previously unseen real homes - making beds, cleaning kitchens, sorting laundry - without per-house fine-tuning. The π-0.5 release also expanded the openpi GitHub repository with new checkpoints, training recipes and evaluation tools. π-0.5 is positioned as the next step toward general-purpose physical intelligence and continues PI's strategy of partial open-source: research weights are public, while production-scale checkpoints and proprietary data remain internal.

Arhitektūra

Flow-matching VLA with co-training for open-world generalisation

π-0.5 builds on the π-0 architecture (PaliGemma 3B multimodal backbone + flow-matching action expert) and introduces a co-training recipe combining heterogeneous data sources: high-quality teleoperation, web-scale image+text, multi-robot cross-embodiment data and large amounts of in-the-wild egocentric video. The training pipeline interleaves robot action prediction with auxiliary objectives (next-token prediction over instructions, captioning, scene reasoning), which significantly improves zero-shot transfer to unseen homes. Inputs remain multi-view RGB + proprioception + natural-language instruction; outputs are flow-matched continuous action chunks. The technical report shows π-0.5 generalising to homes never seen during training, performing multi-step household tasks with natural-language prompts. π-0.5 also improves on long-horizon planning by adopting a hierarchical inference strategy that issues sub-instructions through the same VLA.

Parametri

~3B (PaliGemma backbone + action expert)

Iespējas

• Open-world generalisation across unseen homes
• Long-horizon household tasks: cleaning, bed making, laundry
• Co-training with robot data + egocentric video + web text
• Hierarchical self-prompting for multi-step tasks
• Flow-matching action head for smooth high-frequency control
• Single set of weights deployed across many real homes
• Successor to π-0 with improved language grounding
• Partial open-source via openpi GitHub repository
• Best for: open-world robot research, household manipulation.

Apmācība un licence

Co-trained corpus combining ~10,000+ hours of robot teleoperation, Open-X-Embodiment cross-embodiment data, large-scale egocentric / web video, and web-scale image-text pairs. Multi-stage curriculum mixing action prediction with vision-language objectives.

Licence: Partially open-source via the openpi GitHub repository for research use. Production deployment licensing is handled directly by Physical Intelligence.

Drošības pārbaudes: PI emphasises supervised piloting, force-limited hardware, and controlled household deployment as primary safety mitigations. No public RSP-style framework; ongoing internal evaluation for unsafe-behaviour failure modes.