Generated illustration for Real-Time Action Chunking with Large Models — Paper 13 of 41Seedream 4.0 · GMI Studio

Illustration brief the model wrote

A robot arm holding a component in mid-air, with a ghostly transparent arm showing the next planned grasp. Above the arm, a streaming timeline: one chunk finishing, next chunk starting without gap. No text, no logos, no faces.

Provenance. Brief written by deepseek-ai/DeepSeek-V4-Flash via GMI Cloud on 2026-09-18 from the paper's title, primary link and the editor's notes; reviewed editorially. Graphic rendered by seedream-4-0-250828 through the GMI Studio request queue (request 80e76ccb…), converted to PNG locally. Summaries can be wrong; the linked paper is the source of truth.

Atlas/Physical Intelligence·Physical AI company·2025

Real-Time Action Chunking with Large Models

Model tl;dr

Real-time action chunking for diffusion-based VLAs generates next chunk while previous executes, eliminating discontinuities without training changes.

real-time controlaction chunkingdiffusion policyVLAsmooth executionconcurrent inference

What changed

Model

Previously, using large vision-language-action models for real-time control required either waiting for full inference before acting, causing latency, or using naive smoothing that introduced discontinuities. This paper introduces a method to generate the next action chunk concurrently with executing the previous one, maintaining smooth transitions without any modifications during training.

So what, in plain English

Model

Think of a factory robot that needs to decide its next move while still finishing the current one. This paper lets the robot plan ahead in real-time without jerky pauses, like a chef chopping vegetables while already deciding which pan to reach for next.

Contribution to industry

Model

This technique enables deployment of large VLAs on robots with reasonable latency (100-300 ms), making it practical for industrial settings where cycle times matter. Physical Intelligence likely integrates this into their generalist robot models to allow faster, smoother operation without retraining. It could be used by robotics companies needing real-time action generation with high-capacity models.

Typed judgments

Jev

★★★☆☆ robots · Physical AI 95%

Real hardware 49% · Open code / data 9% · Foundation-model based 95% · Survey / benchmark-only 9%

Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.

Editor's notes

Summary. Real-time chunking lets diffusion- and flow-based VLAs generate the next action chunk while the previous one is still executing, with no training-time changes and without the discontinuities that break naive smoothing.

Industry. Allows remote or on-robot VLA inference with 100-300 ms of latency while keeping industrial cycle times intact as models keep growing.

Read the paper ↗Official research index ↗All 8 Physical Intelligence papers →