Physical Intelligence Pi Model Series

The Physical Intelligence Pi model series is the robot-model sequence discussed in 146. 对Physical Intelligence柯丽一鸣4小时访谈：Pi的开源模型研究，机器人的江湖、族谱与主角 and later extended by 170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人. In K’s account, π0 stands for broad task capability, π0.5 for generalization across new home environments, and π0.6* for task performance improved through Robot Experience Data and Robot Reinforcement Learning.

The later Q2 2026 embodied-intelligence review adds Pi 0.7 as a route where Vision Language Action Models absorb a lightweight future-image or world-model component. Together, the two sources make the Pi sequence a useful case for World Model VLA Fusion: the route is not a clean jump from VLA to world models, but an incremental effort to improve action policies with data, prediction, and evaluation.

Key Points

• π0 is described through tasks such as folding clothes, folding boxes, and table clearing.
• π0.5 is described as a generalization step using data from multiple home-like environments, including Airbnb-style settings.
• π0.6* is described as a performance step where throughput combines speed and successful task completion.
• Pi 0.7, in the later source, adds a lightweight world-model component while remaining close to a VLA route.
• The sequence should be read chronologically: the long K interview was recorded before Pi 0.7 and therefore does not cover it directly.

Connections

• Physical Intelligence and [[KPhysicalIntelligence|K]] — company and researcher explaining the series.
• Robot Generalization Performance Tradeoff — capability, generalization, and performance tension the sequence exposes.
• Robot Experience Data, Robot Reinforcement Learning, and Real Robot Data Strategy — data and learning mechanisms behind π0.6*.
• Vision Language Action Models, World Models, and World Model VLA Fusion — model categories relevant to Pi 0.7.
• Robot Evaluation Problem — why comparing Pi iterations to a single benchmark ladder is difficult.