π0 (pi-zero)

**π0 (pi-zero)** is Physical Intelligence's flagship Vision-Language-Action (VLA) model, announced in October 2024 and described in the paper 'π0: A Vision-Language-Action Flow Model for General Robot Control' (arXiv:2410.24164, authors: Black, Brown, Driess et al.). π0 is the first publicly known 'generalist robot policy' that performs dexterous, long-horizon (multi-minute) manipulation sequences across many different robots — including single- and dual-arm Franka manipulators, ALOHA, UR5, and the Mobile ALOHA humanoid.

The architecture consists of three layers: (1) **VLM backbone** — PaliGemma 3B (Google) as the vision-language foundation, augmented with additional action tokens. (2) **Action expert** — a transformer-style network of ~300M parameters generating a stream of continuous actions (joint velocities) via **flow matching** (instead of action discretization as in RT-2 / OpenVLA — flow matching enables inference in full continuous precision and reduces noise). (3) **Action chunking** — the model predicts 50 future actions simultaneously (a 1-second chunk at 50 Hz), substantially reducing effective latency and producing smooth, coordinated movements.

Training data: ~10,000 hours of demonstrations across **8 robotic platforms** collected by Physical Intelligence + the Open X-Embodiment dataset (1M trajectories). Demonstrations come from teleoperation (VR rig + ALOHA). Tasks in the eval suite: folding T-shirts, packing Amazon-style shipping boxes, clearing a table, loading a dishwasher, folding a cardboard box. **π0-base** (February 2025) was released as open weights under Apache 2.0 (the first 'generalist' weights in history).

Inference: ~1 H100 GPU for full π0 (3B parameters), latency ~50-200 ms per action chunk (50 actions = 1 s of motion). Fine-tuning is available via the `openpi` codebase on GitHub (Apache 2.0). The **π0.5** version (April 2025) introduces co-training with additional multimodal datasets and better out-of-distribution generalization (e.g., unseen rooms).