Pi Series (Physical Intelligence)

Author’s Note

True Star Researcher Team. Model series created by Karol Hausman, Chelsea Finn, Sergey Levine, Pete Florence and others who led Google DeepMind’s RT series.
Exemplary VLA + Teleop Data Approach. Well demonstrates the approach of building Robot Foundation Models with teleoperation-based data collection and VLA architecture.
Open-Source Friendly. Released weights and code via openpi, enabling general users to fine-tune. 1-20 hours of data is sufficient.
Skeptical of Humanoids/Synthetic Data. Focuses on mobile manipulators over humanoids, real data over synthetic data.
Must Follow. A model series worth continuous attention in the Physical AI field.

Pi Series: Robot Olympics Challenge (Physical Intelligence)

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
Self-Improvement via RL: Pi*0.6 uses RECAP methodology to learn from real experience
Fully Open-Source: Complete release of model weights, training code via openpi

Overview

Pi Series is a Vision-Language-Action model series announced by Physical Intelligence starting October 2024.

Item	Details
Company	Physical Intelligence
Founded	2024
GitHub	Physical-Intelligence/openpi
Blog	pi.website/blog

Evolution Timeline

Date	Name	Description	Doc	Official
2024.10.31	Pi0	First Generalist Policy, Flow Matching + PaliGemma 3B	Pi0	Blog
2025.01.16	FAST	5x faster training, DCT + BPE compression	FAST	Research
2025.02.04	openpi	Weights, code released (JAX + PyTorch)	-	Blog
2025.02.26	HIRobot	Human-Robot Interaction research	-	Research
2025.04.22	Pi0.5	Open-World Generalization, Web Data Co-training	Pi0.5	Blog
2025.05.28	Knowledge Insulation	7.5x fewer training steps	-	Research
2025.06.09	Real-Time Chunking	Real-time control in high-latency environments	-	Research
2025.11.17	*Pi0.6**	Self-improvement via RL (RECAP)	Pi*0.6	Blog
2025.12.16	Human-to-Robot	Learning from human videos	-	Research

Model Versions

Version	Released	Key Innovation	Detailed Doc
Pi0	2024.10	Flow Matching VLA, 50Hz control	Pi0
Pi0.5	2025.04	Open-World generalization, Web data Co-training	Pi0.5
*Pi0.6**	2025.11	RL self-improvement, RECAP	Pi*0.6

Core Innovations

Flow Matching Architecture

Alternative to Diffusion, efficiently modeling complex multimodal action distributions. Generates 50Hz action chunks.

FAST Tokenizer

Detailed doc: FAST (Fast Action Tokenizer)

DCT + BPE based action compression achieving 10x compression, 5x faster training.

Knowledge Insulation

Details: pi.website/research/knowledge_insulation

Preserves VLM backbone knowledge while learning robotics. 7.5x fewer training steps via gradient blocking.

Real-Time Chunking (RTC)

Details: pi.website/research/real_time_chunking

Maintains real-time control in high-latency environments (200ms+) using inpainting approach.

RECAP

Detailed doc: Pi*0.6

RL with Experience & Corrections via Advantage-conditioned Policies. Self-improvement through demonstrations + autonomous experience + coaching.

Open Source

openpi Repository

Model	Description
Pi0 base	Pretrained model for fine-tuning
Pi0-FAST base	FAST tokenizer applied version
Pi0 DROID	Franka single arm fine-tuned
Pi0 ALOHA	Bimanual manipulation fine-tuned

JAX: Official implementation
PyTorch: HuggingFace LeRobot integration
Fine-tuning: 1-20 hours data, Consumer GPU capable

Research Publications

Blog Posts

Date	Title
2024.10.31	Pi0: Our First Generalist Policy
2025.02.04	Open Sourcing Pi0
2025.04.22	Pi0.5: Open-World Generalization
2025.11.17	Pi*0.6: Learning from Experience

Research

Date	Title
2025.01.16	FAST: Efficient Robot Action Tokenization
2025.02.26	HIRobot: Interactive Learning
2025.05.28	Knowledge Insulation
2025.06.09	Real-Time Action Chunking
2025.12.16	Human-to-Robot Transfer

Papers

Model	Link
Pi0	arXiv:2410.24164
Pi0.5	arXiv:2504.16054
Pi*0.6	arXiv:2511.14759