Pi Series (Physical Intelligence)

Physical Intelligence's Vision-Language-Action Model Series - Pi0, Pi0.5, Pi*0.6

Pi Series (Physical Intelligence)

Home > Models > Pi Series


Key Significance

  • New Paradigm for VLAs: Flow Matching-based continuous action generation as an alternative to Diffusion/Autoregressive approaches
  • Real Home Environment Generalization: Pi0.5 works in completely new homes never seen during training - new standard for robot generalization
  • Self-Improvement via RL: Pi*0.6 uses RECAP methodology to learn from real experience and continuously improve performance
  • Star Team Startup: Founded by key researchers from Google DeepMind RT series (Karol Hausman, Chelsea Finn, Sergey Levine, Pete Florence)
  • Fully Open-Source: Complete release of model weights, training code, PyTorch/JAX implementation via openpi
  • Revolutionary Efficiency: 5x faster training with FAST tokenizer, 7.5x fewer training steps with Knowledge Insulation
  • 24-Hour Continuous Operation: Pi*0.6 demonstrated espresso making 5:30am-11:30pm, folding 50 new laundry items continuously

Overview

Pi Series is a Vision-Language-Action model series announced by Physical Intelligence starting October 2024. Founded by key researchers who led Google DeepMind’s RT series, they presented a new VLA paradigm based on Flow Matching.

ItemDetails
CompanyPhysical Intelligence
Founded2024
FoundersKarol Hausman, Chelsea Finn, Sergey Levine, Pete Florence, etc.
GitHubPhysical-Intelligence/openpi
Blogpi.website/blog

Evolution Timeline

2024.10 ------- Pi0 ------------------------------------------
                |  First Generalist Policy
                |  Flow Matching + PaliGemma 3B
                |  8 robots, 68 tasks
                |
2025.01 ------- FAST Tokenizer -------------------------------
                |  5x faster training
                |  DCT + BPE compression
                |
2025.02 ------- Open Source (openpi) -------------------------
                |  Weights, code released
                |  JAX + PyTorch (LeRobot)
                |
2025.04 ------- Pi0.5 ----------------------------------------
                |  Open-World Generalization
                |  Works in new homes
                |  Web Data Co-training
                |
2025.05 ------- Knowledge Insulation -------------------------
                |  7.5x fewer training steps
                |  Gradient blocking to preserve VLM knowledge
                |
2025.06 ------- Real-Time Chunking (RTC) ---------------------
                |  Real-time control even in high-latency environments
                |  Inpainting-based approach
                |
2025.11 ------- Pi*0.6 ---------------------------------------
                   Self-improvement via RL
                   RECAP: Learning from experience
                   90%+ success rate, 2x throughput

Model Versions

VersionReleasedKey InnovationDetailed Doc
Pi02024.10Flow Matching VLA, 50Hz controlPi0
Pi0.52025.04Open-World generalization, Web dataPi0.5
Pi*0.62025.11RL self-improvement, RECAPPi*0.6

Core Innovations

1. Flow Matching Architecture

Alternative to Diffusion, efficiently modeling continuous distributions:

FeatureDescription
Continuous DistributionHandles complex multimodal action distributions
High-Frequency ControlGenerates 50Hz action chunks
Transformer IntegrationNatural combination with VLM

2. FAST Tokenizer

Efficiently compresses action sequences:

Raw Actions -> DCT Transform -> BPE Encoding -> 30-60 Tokens
                   |                |
            (JPEG/MP3 method)    (LLM method)
  • 10x compression: Compared to existing tokenization
  • 5x faster training: Compared to Diffusion-based Pi0
  • Dexterous tasks: Enables precise high-frequency control

3. Knowledge Insulation

Preserves VLM’s internet knowledge while learning robotics:

ProblemSolution
Action Expert -> VLM backpropagationGradient Blocking
Robot training damaging language understandingSimultaneous Discrete Action learning
Result7.5x fewer training steps

4. RECAP (RL with Experience & Corrections)

Core of Pi*0.6 - Learning from experience:

+-------------------------------------------------------------+
|                    RECAP Learning Loop                       |
+-------------------------------------------------------------+
|                                                              |
|  +----------+     +----------+     +----------+              |
|  | Demo     | --> | Autonomous| --> | Coaching |             |
|  | (Demo)   |     | (Deploy) |     |(Coaching)|             |
|  +----------+     +----+-----+     +----------+              |
|                        |                                     |
|                        v                                     |
|              +----------------+                              |
|              | Value Function | <- Predicts success prob     |
|              +-------+--------+    per situation             |
|                      |                                       |
|                      v                                       |
|              +-------------------+                           |
|              | Advantage Cond.   | <- Reinforces good actions|
|              +-------------------+                           |
+-------------------------------------------------------------+

5. Real-Time Chunking (RTC)

Real-time control in high-latency environments:

  • Problem: Large VLAs take time for inference, world changes meanwhile
  • Solution: Inpainting approach to maintain executing actions from previous chunk
  • Result: Maintains precision and speed even at high latency

Training Data & Scale

Pi0 Training Data

ItemDetails
Robot Platforms8 (UR5e, Franka, Trossen, ARX, etc.)
Tasks68
External DataOpen X-Embodiment

Pi0.5 Co-training Data

Data TypePurpose
Web DataImage captioning, Visual QA, Object detection
Language DemonstrationsStep-by-step instruction learning
Subtask CommandsHigh-level semantic labels
Robot ActionsMulti-embodiment learning

Performance Highlights

Pi0 vs Baselines

TaskPi0OpenVLAOcto
Bussing (UR5e)97%0%4%
Shirt Folding100%0%0%
Grocery Bagging79%0%0%

Pi*0.6 Improvements

MetricImprovement
Success Rate90%+
Throughput2x or more
Failure Rate2x+ reduction

Real-World Deployment

TaskAchievement
Espresso Making5:30am-11:30pm continuous operation
Laundry Folding50 new items processed continuously
Box Assembly59 chocolate packaging boxes (factory)

Open Source Ecosystem

openpi Repository

ModelDescription
Pi0 basePretrained model for fine-tuning
Pi0-FAST baseFAST tokenizer applied version
Pi0 DROIDFranka single arm fine-tuned
Pi0 ALOHABimanual manipulation fine-tuned
Pi0 LiberoSimulation environment fine-tuned

Framework Support

FrameworkSupport
JAXOfficial implementation
PyTorchHuggingFace LeRobot integration

Fine-tuning Requirements

  • Data: 1-20 hours sufficient
  • Hardware: Consumer GPU capable

Team Background

Previous achievements of Physical Intelligence founding team:

PersonPrevious AffiliationKey Contributions
Karol HausmanGoogle DeepMindRT-1, RT-2 Lead
Chelsea FinnStanford/GoogleMAML, Robotics Transformer
Sergey LevineUC Berkeley/GoogleRL, Robot Learning
Pete FlorenceGoogle DeepMindDense Descriptors

Research Publications

Blog Posts (Chronological)

DateTitleType
2024.10.31Pi0: Our First Generalist PolicyBlog
2025.01.16FAST: Efficient Robot Action TokenizationResearch
2025.02.04Open Sourcing Pi0Blog
2025.04.22Pi0.5: Open-World GeneralizationBlog
2025.05.28Knowledge InsulationResearch
2025.06.09Real-Time Action ChunkingResearch
2025.11.17Pi*0.6: Learning from ExperienceBlog

Papers


Impact

Impact of Pi Series on the robotics field:

  1. Flow Matching Validation: Established as practical alternative to Diffusion
  2. Open-World Standard Setting: New standard for generalization beyond the lab
  3. RL Self-Improvement: Continuous performance improvement even after deployment
  4. Open-Source Ecosystem: Foundation for subsequent research like SmolVLA, LeRobot
  5. Industry Application: 24-hour operation demonstrated in actual factory/home environments

See Also