HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Physical Intelligence—π model familyPhysical Intelligence—π model familyPhysical Intelligence—

π model family

by Physical IntelligenceUS

ResearchStage 1 of 5

π0.7 released April 2026. No commercial product announced yet.

Updated 16 Apr 2026Checked 25 Sep0 updates this week

Milestones

Next · First commercial product
  1. π0 released31 Oct 2024Complete.
  2. π0 open-sourced (openpi)Feb 2025Complete.
  3. π0.5 open-world generalisation22 Apr 2025Complete.
  4. π*0.6 with Recap reinforcement learningNov 2025Complete.
  5. π0.7 steerable generalist model16 Apr 2026Complete.
  6. First commercial productNowCurrent milestone.

Most important updates

  • 16 Apr 2026
  • Nov 2025
  • 22 Apr 2025
  • Feb 2025
  • 31 Oct 2024

Current obstacles

  • No agreed testThere is no agreed test for general robot models. Real-world testing is slow and hard to compare between labs.
  • Speed on the robotModels with billions of parameters must answer in tens of milliseconds, which takes clever tricks.

Physics limits

  • Robot data is scarce and slow to collectπ0 was trained on roughly 10,000 hours of robot data, while language models read trillions of words. Robot data is recorded in real time with real hardware, so it grows slowly and costs a lot.
  • Contact physics is hard to simulateFriction, soft objects and cloth behave chaotically: tiny differences change the outcome. Simulated practice therefore transfers imperfectly, and real-world data stays essential.
  • Errors compound over long tasksA task of 20 steps at 95% success each finishes only about 36% of the time. Useful home and factory work needs near-perfect steps or the ability to notice and fix mistakes.

How it works

3 parts
A lab robot arm on a wheeled base, not Physical Intelligence's: π models turn camera images and a spoken task into motions for robots like this
A lab robot arm on a wheeled base, not Physical Intelligence's: π models turn camera images and a spoken task into motions for robots like thisPhoto: Dzikra muhammad Imtiyaz · CC BY-SA 4.0 (opens commons.wikimedia.org)
Perceive

Image-and-language backbone

A pretrained AI that already knows images and words reads the cameras and the task.

Act

Action module

A separate part creates a short chunk of future motion by refining random noise step by step, giving smooth, precise moves.

Improve

Learning from deployment

Since π*0.6, the model also trains on its own attempts and human corrections, favouring moves that helped.

Spec sheet

Specπ model family
π0 VLM backbone~3B parameters (PaliGemma)R (reported)
π0 action expert~300M parametersR (reported)
Action rateUp to 50 Hz action chunksR (reported)

R reported by the company

Papers & demos

6 items
  1. 16 Apr 2026paper
    π0.7: a steerable generalist robotic foundation model with emergent capabilities (opens arxiv.org)Early signs of combining skills, like folding laundry on a robot never trained on laundry.
  2. Nov 2025paper
    π*0.6: a VLA that learns from experience (Recap) (opens arxiv.org)Learning from real deployment data made it faster and more robust than copying demos alone.
  3. 22 Apr 2025paper
    π0.5: a VLA with open-world generalization (opens arxiv.org)Showed mobile robots cleaning homes that were not in the training data.
  4. Feb 2025code
    openpi: open-source π0 weights and code (opens github.com)Made a top robot model free for anyone to use and build on.
  5. Jan 2025paper
    FAST: efficient action tokenization for VLA models (opens arxiv.org)A new way to compress robot motions into tokens that made some robot models train much faster.
  6. 31 Oct 2024paper
    π0: a vision-language-action flow model for general robot control (opens arxiv.org)Set the recipe: an image-and-language AI plus an action module, for skilled control of many robot types.

Update log

5 updates

Thu 16 Apr

  • Major: BlogResearch

Nov 2025

  • Major: BlogResearch

Tue 22 Apr 2025

  • Major: BlogResearch

Feb 2025

  • Major: BlogSoftware

Thu 31 Oct 2024

  • Major: BlogResearch

About Physical Intelligence

The team behind π model family

Physical Intelligence

Robot brains (π0, π0.5, π0.7)

Physical Intelligence (π) is a San Francisco lab building AI brains that work on many kinds of robots. Founded in 2024 by researchers from Google DeepMind, Stanford and UC Berkeley, it shares some models openly but sells nothing yet.

  • Sets the research pace: π0 (2024), π0.5 (2025), π*0.6 (2025) and π0.7 (Apr 2026).
  • Valued at about $11B in 2026, roughly double 4 months earlier, with no revenue.
  • Its open weights (openpi) are the default starting point for many robot labs.

pi.website (opens pi.website)Company profile →

Founded
20242 yrs
Headquarters
United States
Status
Private
Valuation
$11BprivateJan 2026
Raised
$2.1B4 rounds
Last round
Growth round · $1B2026$11B post
Works in
RoboticsRobot brainsAI
Coverage
1 program · 7 updateslatest 16 Apr 2026checked 25 Sep
Lead investors
Founders FundCapitalGThrive Capital
Investors
Thrive Capital (opens thrivecap.com)Lux Capital (opens luxcapital.com)Founders Fund (opens foundersfund.com)CapitalG (opens capitalg.com)
People
Karol Hausman (opens karolhausman.github.io)Co-founder & CEOSergey Levine (opens en.wikipedia.org)Co-founderChelsea Finn (opens en.wikipedia.org)Co-founder
Primer