pi-zero

Summary

π0 (pi-zero) is Physical Intelligence's open-source vision-language-action robot foundation model — a VLM backbone with a flow-matching action head producing 50 Hz trajectories. Trained on 10,000+ hours across 7 robotic platforms and 68 tasks, it has seeded a major open-source robotics ecosystem.

Overview

π0 (pi-zero) is Physical Intelligence's first-generation robot foundation model — a vision-language-action (VLA) flow model designed to control diverse robot embodiments across diverse tasks using a single shared model. Introduced in late 2024, π0 was trained on 10,000+ hours of robot data spanning 7 robotic platforms and 68 unique tasks, and demonstrates strong zero-shot and fine-tuned performance on complex real-world tasks including laundry folding, table bussing, grocery bagging, box assembly, and object retrieval. The model was open-sourced in 2025, making it one of the most consequential open-weight robotics releases to date.

π0's architecture inherits semantic knowledge and visual understanding from internet-scale pretraining by starting from a pre-trained vision-language model (VLM), then adds an action-generation component using flow matching to produce smooth, real-time action trajectories at 50 Hz. This architectural choice — VLM backbone + flow-matching action head — has become a widely-cited template in robotics foundation model research. A subsequent π0.5 variant released in September 2025 extended π0 with better open-world generalization, and π0-FAST was introduced as a higher-frequency variant.

Specifications

  • Developer: Physical Intelligence
  • Release Date: Late 2024 (π0 base); 2025 (open-source release); September 2025 (π0.5)
  • Type: Vision-Language-Action (VLA) flow model for robot control
  • Architecture: Pre-trained VLM backbone + flow-matching action generation head; produces actions at 50 Hz
  • Training Data: 10,000+ hours of robot data across 7 robotic platforms and 68 unique tasks
  • License: Open-source via openpi GitHub repo
  • Access: Hugging Face + openpi GitHub + Physical Intelligence direct
  • Related Models: π0.5 (improved open-world generalization, September 2025); π0-FAST (higher-frequency variant)

Capabilities

Cross-Embodiment Generalization (defining capability): A single model that controls 7 distinct robotic platforms — a significant departure from platform-specific policies that have dominated prior robotics research. Fine-tuning to a new platform requires only 1–20 hours of additional data.

Real-World Task Performance: Demonstrated on complex real-world tasks including laundry folding, table bussing, grocery bagging, box assembly, and object retrieval — meaningful demonstrations beyond the simple pick-and-place tasks that have historically dominated robotics demos.

50 Hz Real-Time Action Generation: Flow-matching action head produces smooth, real-time action trajectories at 50 Hz — sufficient frequency for natural manipulation tasks without separate motion-planning systems.

VLM-Inherited Semantic Knowledge: Starting from a pre-trained vision-language model lets π0 leverage internet-scale visual and semantic understanding for tasks involving object recognition, language instructions, and contextual reasoning.

Zero-Shot and Fine-Tuned Modes: π0 demonstrates strong zero-shot performance on tasks within its training distribution and rapid fine-tuning (1–20 hours of data) for tasks outside it.

Open-Source Distribution: Available via openpi GitHub repo and Hugging Face — has seeded an academic and industrial ecosystem of fine-tunes and downstream applications.

Limitations

Fine-Tune Data Required for New Tasks: While 1–20 hours of fine-tune data is much less than prior task-specific systems required, it still represents a meaningful data collection effort for each new deployment scenario.

Dexterity Limits: π0's manipulation capabilities are strong for the demonstrated task list but do not match human dexterity on the most complex bimanual tasks. The π0.5 release explicitly addresses some open-world generalization gaps.

Hardware Constraints: π0's training and deployment require GPU compute meaningful enough to run a vision-language-action model in real-time — practical for research labs and commercial deployments, but not yet for very low-resource embedded systems.

No Skill-Library Integration: π0 is a single end-to-end policy; integration with separate skill-library systems, motion planners, or task-decomposition frameworks requires custom integration work.

Recent Developments

  • Open-Source Release (2025): π0 weights released via the openpi GitHub repo — one of the most consequential open-weight robotics releases to date. Has seeded an academic and industrial ecosystem of fine-tunes and downstream applications analogous to Stable Diffusion's role in image generation.
  • π0.5 Release (September 2025): Upgraded variant with improved open-world generalization — better handling of tasks and environments outside the original training distribution.
  • π0-FAST Variant: Higher-frequency variant for applications requiring tighter control loops.
  • Industry Position: π0 represents the leading independent open-weight robotics foundation model. Skild AI ($14B valuation) operates in the same category but with a closed, commercial distribution model. Competitors at Google DeepMind, NVIDIA, and the in-house teams at humanoid hardware companies (1X, Figure) operate at similar scale but with proprietary distribution.

Last Updated

May 7, 2026

→ Back to Models