VLA Gallery / All models / π family

Real-Time Chunking (추론시 RTC)

Physical Intelligence, UC Berkeley · 2025-06

기타π familyN/Aflow matching

Inference technique applied to diffusion/flow VLAs without retraining; inpaints the next chunk to match the actions being executed, so execution continues smoothly despite inference latency.

Paper

Real-Time Execution of Action Chunking Flow PoliciesarXiv 2506.07339

Overview

Category

π family

Type

기타

Released

2025-06

Organization

Physical Intelligence, UC Berkeley

Open weights

No

Parameters

N/A (uses π0.5)

Architecture

Backbone LLM

Gemma 2B (applied to π0.5)

Vision encoder

SigLIP 400M (π0.5)

Action head

π0.5 flow expert + inpainting chunk generation (no retraining) (flow matching)

Chunk

50 steps

Control rate

50 Hz

Architecture (original summary)

Applied to π0.5 (H=50, Δt=20 ms); generates the next chunk by inpainting conditioned on already-committed actions (no retraining)

Training

Pretraining accelerators

N/A

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

No VLA retraining. Kinetix sim policy (4-layer MLP-Mixer, 32 epochs) about 1.5 h on H100×2 per environment, whole experiment on 8 or fewer H100

Data type

N/A

Data scale

N/A

Main datasets

N/A

Tasks & evaluation

6 bimanual tasks (lighting a candle, plugging Ethernet, making a bed (mobile), shirt folding, batch folding, dishwashing (mobile)), 10 trials per task·method, 480 episodes·28 h total; 12 Kinetix dynamic tasks

Robot

Bimanual·mobile bimanual manipulators (π0.5 real-robot platforms)

Inference

Inference device & speed

RTX 4090, bf16, 5 denoising steps: baseline 76 ms vs RTC 97 ms; π0 KV cache prefill 46 ms

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Local PC

Reinforcement learning & base

Reinforcement learning

Not used · Inference technique (no retraining)

Base control

Wheeled (Format not stated) · RTC applied to 2 π0.5 mobile bimanual tasks (Make bed·Dishes in sink); whether base velocity is included not stated in paper (π0.5 baseline)

Notes

Key features

Inference technique applied to diffusion/flow VLAs without retraining; inpaints the next chunk to match the actions being executed, so execution continues smoothly despite inference latency. Evaluated on 6 bimanual tasks, 480 episodes, 28 hours of robot execution.

Notes

An inference method that does not change training, so weights and data fields are N/A. Training-time RTC (2025-12) is a separate row.

Verification

Corrections applied