Physical IntelligenceNot yet claimed

About Physical Intelligence

Physical Intelligence is a San Francisco-based company bringing general-purpose AI into the physical world. As its official website puts it, the company develops learning algorithms and foundation models intended to control any robot to do any task, powering the robots of today and the physically actuated devices of the future rather than building robot hardware itself. Its model family is the π (pi) series of vision-language-action models, beginning with π0, its first generalist policy, released in October 2024 and open-sourced with weights and code in February 2025, and continuing through π0.5, which extended the policy to open-world generalisation — including controlling a mobile manipulator to clean up an entirely new kitchen or bedroom — to later steerable models with memory and online reinforcement-learning methods for precise, longer tasks.

In October 2026 the company made its latest robotics models available through the Pi API beta, through which users can upload data, fine-tune models and run cloud-hosted inference on their own robots, and it publishes partner work showing its models applied to real-world problems. Its research output, published openly on its own site, covers robot action tokenisation, real-time action chunking, human-to-robot transfer and models that learn from experience. Founded in 2024, Physical Intelligence is a model-layer company: robot makers and operators supply the hardware, and its policies supply the generalist control.

Products & services

• π0 — first generalist robot policy (vision-language-action model); weights and code open-sourced in February 2025
• π0.5 and later π-series models — generalist policies with open-world generalisation, memory and steerable behaviour
• Pi API (beta, from October 2026) — data upload, model fine-tuning and cloud-hosted inference for customers' own robots
• FAST robot action tokenisation and related published research methods for training generalist policies

Questions partners ask

GlobleMatch Trust Signals

A directory that doesn’t sell stars. Trust Signals are built only from verifiable facts — never from anonymous ratings, and never for sale. There is no single hidden score.

Public information coverage	8 of 9 fields on the standard profile could be filled from public sources. This counts how much public information exists about the company — not how good the company is. The nine fields are listed in the methodology. Claiming the profile lets the company complete the remaining fields itself.
Last verified	6 November 2026 — facts on this profile were checked against the sources listed below. Profiles are re-checked on a published 90-day cycle; this date is never quietly refreshed.

Claim this profile to unlock response record, directory tenure and peer endorsements. The full methodology — what each signal measures and how it is calculated — is published in the Trust Signals methodology, including a change log.

Sources & verification

• Company website — physicalintelligence.company (official site: profile facts, headquarters, markets and product range). Checked 6 November 2026.
• Figures described in this profile as company-stated come from the company’s own public materials and were not independently audited by GlobleMatch.

Facts last verified: 6 November 2026. Something wrong? A listed company can request a correction or removal at any time; verified requests are actioned.

Related companies

More in Robotics & hardware: Mujin (mujinos platform, Japan) · Field AI (foundation models for general-purpose robots in unstructured environments, United States) · MMI (Medical Microinstruments) (robotic microsurgery and supermicrosurgery systems, United States)

Robotics on GlobleMatch: the GlobleMatch Robotics 100 programme tracks leading robotics companies worldwide, and claimed directory profiles can be considered for its seats.