VLA Gallery / All models / π family

π*0.6 (RECAP)

Physical Intelligence · 2025-11

VLAπ familyTotal not disclosedhybrid

Iteratively trains on demonstrations·autonomous rollouts·expert corrections with advantage-conditioned offline RL (RECAP).

Paper

π*0.6: a VLA That Learns From ExperiencearXiv 2511.14759

Overview

Category

π family

Type

VLA

Released

2025-11

Organization

Physical Intelligence

Open weights

No

Parameters

Total not disclosed; policy = π0.6 (Gemma 3 4B + 860M); value function uses a 670M VLM backbone

Architecture

Backbone LLM

Gemma 3 4B

Vision encoder

SigLIP 400M

Action head

π0.6 action expert + value function (670M backbone) (hybrid)

Chunk

Not disclosed

Control rate

50 Hz

Architecture (original summary)

π0.6 policy (Gemma 3 4B + 860M flow action expert, joints·gripper 50 Hz) + distributional value function with Gemma 3-initialized 670M backbone (201 bins); 'Advantage: positive/negative' text conditioning

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed (both value-function training·RECAP iteration compute)

Data type

Real-robot teleoperation demonstrations + autonomous rollouts + teleoperation corrections

Data scale

Tens of thousands of hours of pretraining demonstrations; per-iteration examples: laundry 300 trajectories (4 robots), diverse laundry eval 450·correction 287, failure removal autonomous about 1000·correction 280+378 (3 robots), box demonstrations 600·correction 360, espresso correction 429·autonomous 414

Main datasets

PI pretraining mix (π0.5 recipe) + per-task RECAP-collected data

Tasks & evaluation

Laundry (T-shirts·shorts, 200 s), 11 types of diverse laundry (500 s), failure removal (orange T-shirt, 200 s), espresso (200 s), box assembly (600 s); 90%+ success except diverse laundry; espresso 13 hours continuously, laundry in a new home 2+ hours without interruption

Robot

Static bimanual (two 6-DoF arms, parallel grippers, 3 cameras, 50 Hz joint control)

Inference

Inference device & speed

Not disclosed in paper; 63 ms on 1× H100 per π0.6 model card

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Used (VLA RL post-training) · RECAP: after pretraining with offline RL, trains a value function (670M) on demos·autonomous rollouts·human correction data and iteratively improves the policy via advantage conditioning

RL evidence

https://arxiv.org/html/2511.14759v1

Base control

Fixed · No base control (static bimanual, joint positions 50 Hz)

Notes

Key features

Iteratively trains on demonstrations·autonomous rollouts·expert corrections with advantage-conditioned offline RL (RECAP). More than doubled successes per hour on hard tasks and cut failure rate by about half.

Notes

Compute for all stages Not disclosed. The 2×+ throughput is for diverse laundry·espresso. Boxes assembled for real factory packaging.

Verification

Confirmed