VLA Gallery / All models / π family

π0.5

Physical Intelligence · 2025-04

VLAπ familyTotal not disclosedflow matching

One model first predicts subtask text and then outputs actions, co-trained with web, multi-robot, and subtask-annotated data.

Paper

π0.5: a Vision-Language-Action Model with Open-World GeneralizationarXiv 2504.16054

Overview

Category

π family

Type

VLA

Released

2025-04

Organization

Physical Intelligence

Open weights

Partial

Parameters

Total not disclosed; VLM 2B (Appendix E), Fig. 3 'SigLIP (400M) + Gemma (2B)' (pretraining†)/'SigLIP (400M) + Gemma (2.6B)' (post-training†); action expert 300M

Architecture

Backbone LLM

Gemma 2B · PaliGemma config 2B VLM

Vision encoder

SigLIP 400M

Action head

flow matching expert 300M (FAST tokens in pretraining) (flow matching)

Chunk

50 steps

Control rate

50 Hz

Architecture (original summary)

PaliGemma-initialized 2B VLM + 300M action expert; pretraining on FAST discrete tokens, flow matching continuous chunks added in post-training (horizon 50, 50 Hz, 10 denoising steps, α=10.0)

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Discrete tokens 280k steps + post-training 80k steps / batch not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed in paper. openpi pi05_full_droid_finetune: H100×8 about 2 days, 100k steps·batch 256 (about 1 epoch), about 5 days for 240k iter from PaliGemma; pi05_libero 30k steps·batch 256 (time Not disclosed). LeRobot: 1× 80 GB GPU, batch 64·30k steps (secondary)

Data type

Real-robot teleoperation + web vision-language + language-instructed demonstrations·subtask annotations

Data scale

Mobile manipulation about 400 h·about 100 homes; 97.6% of stage-1 training examples are non-mobile data; language instruction (VI) demonstrations are about 11% of high-level mobile examples; openpi pi05_base has 10k+ h of robot data

Main datasets

PI's own MM·ME·CE data, OXE; CapsFusion, COCO, Cambrian-7M, PixMo, VQAv2

Tasks & evaluation

3 real homes unseen in training + mock homes; kitchen and bedroom tidying (dishes into sink, drawer organizing, laundry basket, bed making), 10~15-minute tasks, 10 trials per task; scaling over 3, 12, 22, 53, 82, 104 training locations; openpi LIBERO 98.8/98.2/98.0/92.4, mean 96.85

Robot

Home mobile manipulator (arm·gripper·torso lift target pose + base velocity, 50 Hz) + non-mobile·cross-embodiment robots

Inference

Inference device & speed

Not disclosed in paper; RTC paper: RTX 4090, model latency 76 ms with 5 denoising steps

Inference memory

> 8 GB (common to openpi README)

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Joint supervised learning on web·multi-robot·subtask data

Base control

Wheeled (Direct base velocity/pose output) · Policy directly outputs base target velocity (2D linear, 1D angular) as 50 Hz chunks (end-to-end manipulation and navigation)

Notes

Key features

One model first predicts subtask text and then outputs actions, co-trained with web, multi-robot, and subtask-annotated data. Performed 10~15-minute tidying tasks in 3 real homes unseen in training.

Notes

Data hours (h) is the '10k+ hours' (lower bound) for pi05_base in the openpi README; the paper states only about 400 h for the mobile portion. Appendix E num_heads=18 differs from 8 in openpi gemma_2b. Appendix 'action horizon of 50, i.e. H = 49' is quoted verbatim. openpi π0.5 (released 2025-09) was trained with KI, so it may differ from the paper recipe, and supports only the flow head.

Verification

Confirmed