Physical Intelligence

A Stanford assistant professor in computer science and robotics, Chelsea Finn earned her PhD at MIT under Leslie Kaelbling, where she focused on learning for robotic manipulation and reinforcement learning. Prior to founding Physical Intelligence, she was a postdoctoral researcher at Google Brain/DeepMind, contributing to foundational work in meta-learning — most notably MAML (Model-Agnostic Meta-Learning) — and offline imitation learning from diverse demonstration data. That thread of research directly informs Physical Intelligence's thesis: that large, general-purpose vision-language-action models can generalize across robot embodiments. She also leads Stanford's IRIS Lab, which produces recurrent work on robot foundation models and scalable imitation learning.

Earned a PhD in computer science from the University of Southern California, working on robot learning and reinforcement learning with Stefan Schaal. Held a postdoctoral position at Stanford University, then joined Google Brain / Google Robotics as a research scientist, where he contributed to large-scale robot learning systems including the RT-1 and RT-2 foundation-model lines. In 2024, co-founded Physical Intelligence (pi.ai), a San Francisco lab building general-purpose vision-language-action (VLA) foundation models capable of controlling diverse robot hardware. The company's π0 model and subsequent π0.5 represent a bet that a single neural architecture can generalize across tasks and robot embodiments — a premise that contrasts with more task-specific approaches prevalent in industrial robotics.

An early Stripe employee (reportedly among the first handful of hires), spent roughly six years there helping build core payments infrastructure and later leading Stripe Issuing. After departing, founded a namesake venture fund that placed early bets on companies including Figma, Ramp, and Deel — establishing a reputation for sharp, pattern-recognition-driven investing before later turning operator again. At Physical Intelligence, brings that financial and infrastructure-building background to the problem of commercializing general-purpose robot foundation models. Publicly known for a low-key, data-obsessed posture: has published detailed quantitative breakdowns of startup unit economics and is frequently described as operating with unusual analytical rigor for a founder-investor hybrid.

An associate professor at UC Berkeley in EECS, where he leads the Robot Learning Lab within BAIR. He earned his PhD at Stanford (2014), advised by Daphne Koller, and spent time as a research scientist at Google, developing large-scale reinforcement learning and robot manipulation systems. That body of work — spanning model-based RL, imitation learning, and vision-based control — directly informed the architecture of Physical Intelligence's general-purpose vision-language-action models. He is known for an exacting, empirical approach to robotics research and a prolific publication record that has helped define the modern learning-for-robotics paradigm.