VLA Gallery / All models / π family

π0

Physical Intelligence · 2024-10

VLAπ family3.3Bflow matching

The first π-family model, attaching a separate flow matching action expert to a VLM backbone, pretrained on about 10,000 h of data from 7 robot configurations·68 tasks.

Paper

π0: A Vision-Language-Action Flow Model for General Robot ControlarXiv 2410.24164

Overview

Category

π family

Type

VLA

Released

2024-10

Organization

Physical Intelligence

Open weights

Yes

Parameters

3.3B (VLM 3B + action expert 300M); π0-small 470M for comparison

Architecture

Backbone LLM

Gemma 2B · PaliGemma 3B config

Vision encoder

SigLIP 400M

Action head

flow matching action expert 300M (10 steps) (flow matching)

Chunk

50 steps

Control rate

Up to 50 Hz

Architecture (original summary)

PaliGemma 3B VLM + 300M action expert (trained from scratch); flow matching continuous action chunk H=50, up to 50 Hz, 10 integration steps

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

700k steps (compute-parity variant 160k) / batch not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed in paper (post-training data 5~100+ h per task). openpi: LoRA >22.5 GB (RTX 4090), Full >70 GB (A100 80GB/H100), pi0_libero 30k steps·batch 32 (time Not disclosed). LoRA A800×4 about 7 h·30k steps·batch 32 (secondary, retrained in X-VLA paper App. K)

Data type

Real-robot teleoperation (in-house + open data)

Data scale

About 10,000 h; in-house 903M timesteps; 68 tasks; open data share 9.1%

Main datasets

PI's own data, OXE (Magic Soup), Bridge v2, DROID

Tasks & evaluation

Shirt folding, Bussing easy/hard, Grocery bagging, Toast; multi-stage: laundry folding, mobile laundry, mobile dryer, dryer unloading, Table bussing, box building, To-go box, egg packing; 5 fine-tuning tasks; openpi LIBERO 96.8/98.8/95.8/85.2, average 94.15

Robot

UR5e, bimanual UR5e, Franka, bimanual Trossen, bimanual ARX·AgileX, mobile Trossen·ARX, mobile Fibocom (7 configurations)

Inference

Inference device & speed

Onboard RTX 4090 73 ms (image encoder 14 + observation 32 + action×10 27 ms), remote 86 ms; openpi inference >8 GB

Inference memory

> 8 GB (openpi README)

Weights bf16 (calc.)

6.6 GB

Where it runs

Onboard or remote PC

Reinforcement learning & base

Reinforcement learning

Not used · flow matching imitation learning

Base control

Wheeled (Direct base velocity/pose output) · In the mobile configuration, base actions are part of the policy output (Trossen·ARX 2-dim, Fibocom 3-dim); otherwise fixed arm

Notes

Key features

The first π-family model, attaching a separate flow matching action expert to a VLM backbone, pretrained on about 10,000 h of data from 7 robot configurations·68 tasks. Performs multi-stage long-horizon tasks such as laundry folding·box assembly.

Notes

Task-robot combinations weighted by n^0.43. Current arXiv PDF is v4 (2026-01-08); latency tables etc. are also in v1. PI blog unverified (403). openpi weights released since first commit 2025-02-03 (Apache-2.0).

Verification

Confirmed