Physical Intelligence

Physical Intelligence and the pi0 Model

Physical Intelligence

Home > Companies > Physical Intelligence


Overview

Physical Intelligence (π) is a robotics AI startup founded in 2024, developing general-purpose robot foundation models. Founded by a team from Google DeepMind, OpenAI, and Stanford, the company raised the largest seed funding round in robotics history.

ItemDetails
HeadquartersSan Francisco, CA
FoundedMarch 2024
CEOKarol Hausman (formerly Google DeepMind)
Funding$400M+ (Seed: $70M, Series A: $400M)
Valuation$2.4B (as of November 2024)

Founding Team

Co-Founders

NamePrevious RolePosition
Karol HausmanGoogle DeepMind (RT-2)CEO
Sergey LevineUC Berkeley (RL Expert)Chief Scientist
Chelsea FinnStanford (MAML, Meta-learning)Research
Brian IchterGoogle DeepMind (RT-2)Research

Key Investors

  • Thrive Capital, Lux Capital
  • Khosla Ventures, OpenAI
  • Jeff Bezos, Sequoia

pi0 Model

Key Features

ItemDetails
Parameters3.3B
ArchitecturePaliGemma + Flow Matching
Core TechnologyAction Expert with Flow Matching
Open SourceReleased in 2025

Flow Matching Approach

Uses Flow Matching instead of Diffusion:

[Noise] ──Flow Matching──→ [Action Chunk]
        (faster inference)
  • Faster inference than Diffusion
  • Well-suited for continuous action space
  • Capable of learning multi-modal action distributions

Performance

  • Trained on 51 tasks across 20 robot configurations
  • Zero-shot generalization capability
  • Superior performance compared to single-robot policies

Data Collection

Diverse Robot Platforms

  • Single-arm robots
  • Bimanual arms
  • Humanoid upper body
  • Mobile manipulators

Data Characteristics

  • Cross-embodiment data
  • Diverse environments (homes, warehouses, offices)
  • Includes dexterous manipulation

Approach

”Bring GPT to Robotics”

LLM Success = Large-scale Data + Transformer + Scaling

Physical Intelligence Goal = Apply same approach to robotics

Three Core Principles

  1. Scaling: More data, larger models
  2. Generality: Not limited to specific robots/tasks
  3. Foundation Model: Fast adaptation through fine-tuning

Roadmap

TimelineMilestone
2024.03Company founded
2024.10pi0 release
2024.11$400M Series A
2025pi0 open source release
2025+Commercial deployment

References


See Also