Physical Intelligence

What they built

Physical Intelligence (PI, stylized as π) is a foundation-model company for robotics. The flagship products are π0 and π-0.5 — transformer-based models trained on cross-embodiment robot data that enable a wide range of manipulation, mobility, and dexterous tasks when deployed on diverse robot hardware. The models are released open-weight (code and weights available), positioning Physical Intelligence as the central open ecosystem for robot learning, similar to Meta’s Llama strategy in general AI. Commercial licensing and enterprise partnerships layer on top of the open release.

How they got here

Physical Intelligence was founded in 2024 by Karol Hausman, Chelsea Finn, Sergey Levine, Brian Ichter, and Quan Vuong — a group that represents much of the modern robotics-research world. Levine (UC Berkeley) and Finn (Stanford) are among the most-cited researchers in the field; Hausman, Ichter, and Vuong came from Google Brain and DeepMind’s robotics groups. The founding thesis was explicit: robots would reach human-level manipulation and mobility via foundation-model scaling applied to robotic data, and the company best positioned to do that was one dedicated solely to the research, not a robot-hardware company doing AI on the side.

The capital raised at a rate few research-first companies have matched. A $400M Series A in 2024 was followed by a $600M Series B in November 2025 led by CapitalG (Alphabet’s growth fund), pricing Physical Intelligence at $5.6B. By March 2026, the company was reportedly in talks for another $1B round at $11B+ — doubling the valuation in four months. The open-release strategy for π0 in February 2026 (weights + code) was a significant ecosystem move, attracting robotics developers who had previously relied on closed, vendor-specific ML stacks.

What’s ahead

Three questions define Physical Intelligence’s trajectory. First, commercial model: open-weight releases don’t pay bills directly, so the company needs to build a licensing, managed-inference, or enterprise-platform revenue layer without fragmenting the ecosystem. Second, capability progression: π-1 and beyond need to continue closing the gap to human-level dexterous manipulation; the category is not yet mature enough that any model works reliably in arbitrary environments. Third, humanoid integration: if Figure, 1X, Apptronik, and others adopt Physical Intelligence models as their cognitive backbone, the company wins the category. If they build in-house, Physical Intelligence’s addressable market narrows meaningfully.

Why it matters

Physical Intelligence is the most important AI research-first company of the physical-AI era — the one whose trajectory will determine whether robotics foundation models become a winner-take-most category or fragment across hardware companies. For founders in robotics and physical-AI categories, PI’s open-weight strategy matters even if you’re a competitor: you’ll either build on their models or build against them. For investors, Physical Intelligence is a pure-play exposure to the foundation-model-of-robots thesis, with a research team whose academic credentials are unmatched in the space.