VLA Gallery / All models / π family

π0.6

Physical Intelligence · 2025-11

VLAπ familyTotal not disclosedhybrid

Scaled the backbone to Gemma 3 4B and the action expert to about 860M, and added rich metadata conditioning.

Paper

No paper · Related π*0.6: a VLA That Learns From Experience

Overview

Category

π family

Type

VLA

Released

2025-11

Organization

Physical Intelligence

Open weights

No

Parameters

Total not disclosed; Gemma 3 4B backbone + about 860M action expert

Architecture

Backbone LLM

Gemma 3 4B

Vision encoder

SigLIP 400M

Action head

flow expert about 860M + FAST tokens (KI) (hybrid)

Chunk

Not disclosed

Control rate

50 Hz

Architecture (original summary)

Gemma 3 4B VLM (figure: SigLIP 400M + Gemma 4B) + about 860M action expert (same layer count as the backbone); KI trains flow matching continuous chunks and FAST discrete tokens simultaneously; up to 4 images of 448×448

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

N/A (evaluated without per-task fine-tuning)

Data type

Real-robot teleoperation + web vision-language + subtask prediction data

Data scale

Time not stated on the card; the π*0.6 paper describes the pretraining data as 'tens of thousands of hours of demonstrations'

Main datasets

Inherits the π0.5 mixture (in-house·external cross-embodiment, home mobile·non-mobile, web multimodal)

Tasks & evaluation

Static: shirt folding, laundry folding, box assembly, Table bussing; mobile: laundry pickup, bed making, dishwashing, putting items in drawers; language following and new-skill generalization; full box assembly 20%; metrics are success rate, progress, and successes per hour

Robot

Static bimanual + mobile manipulator (base camera, up to 2 wrist cameras, rear camera for mobile)

Inference

Inference device & speed

Single H100, 63 ms per chunk with 5 denoising steps·3 cameras

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Supervised learning (KI). The RL version is π*0.6

Base control

Wheeled (Format not stated) · Includes mobile bimanual tasks (laundry basket·bed making·dishes to sink, etc.); base action format Not stated in model card

Notes

Key features

Scaled the backbone to Gemma 3 4B and the action expert to about 860M, and added rich metadata conditioning. Without per-task fine-tuning, it folds laundry reliably and fully assembles boxes 20% of the time.

Notes

The model card (2025-11-17) original reads 'fold laundry reliably'. openpi has no π0.6 checkpoint as of the 2026-08 HEAD. A summed total of components is not in the sources, so it is not given.

Verification

Confirmed