π0.5 (pi-zero-5)

π0.5 (read: pi-zero-five, or pi-zero-0.5) is a second-generation VLA (Vision-Language-Action) model developed by Physical Intelligence (π). It is an evolution of the original π0, announced in 2024. While π0 focused on demonstrating the capabilities of a generalist robot control model with pre-training and rapid fine-tuning on specific tasks, π0.5 targets open-world manipulation — the ability to perform household tasks (cleaning rooms, tidying objects, working in kitchens) in diverse, previously unseen homes, without additional fine-tuning.

Architecturally, π0.5 is built on the same foundation as π0 — a flow matching head on top of a large VLM (Paili/Gemma-based) with tokenized actions. The key innovation is the massive scale and diversity of training data: tens of thousands of hours of teleoperated demonstrations collected in many different homes by Physical Intelligence operators. The data covers hundreds of different everyday objects, various surfaces, lighting conditions, and room configurations, forcing true generalisation rather than just within-domain interpolation.

π0.5 is evaluated in zero-shot and few-shot mode on cleaning, tidying and kitchen tasks in different homes. Results indicate significantly better generalisation than previous imitation learning models and competitive benchmarks (RT-2, OpenVLA, Octo). Physical Intelligence plans further scale (π1, π2) on the same trajectory.

The model is de facto closed (proprietary weights), although the architecture and methodology are described in a technical report. The inference framework openpi is partially available open-source. The system is deployed on hardware with RGB or RGB-D cameras and standard dexterity manipulators.