← Back ICRA 2026

Closed-Loop Action Chunks with Dynamic Corrections for Training-Free Diffusion Policy

Pengyuan Wu, Pingrui Zhang, Zhigang Wang, Dong Wang, Bin Zhao, Xuelong Li

PDF

AI summary

Key figure (auto-extracted from paper)

DCDP enables real-time closed-loop correction for diffusion policies without retraining, boosting adaptability in dynamic scenarios with minimal computational overhead.

Diffusion Policy Closed-Loop Control Dynamic Manipulation Action Chunking Training-Free Adaptation Robotic Policy

Problem

Existing diffusion policies generate action chunks in an open-loop manner, lacking the ability to rapidly adapt to dynamic environmental changes or moving targets during execution.

Approach

The framework injects a lightweight, self-supervised dynamic feature encoder and cross-attention module into the inference stage to continuously extract environmental dynamics and correct action chunks in real-time without retraining the base policy.

Key results

19% adaptability improvement on dynamic PushT simulations
5% additional computational overhead
Training-free plug-and-play integration
Real-time closed-loop correction without retraining

Why it matters

Provides a modular, training-free solution for robotic manipulation that balances long-horizon planning with immediate reactivity, benefiting researchers and practitioners working with diffusion-based policies in dynamic environments.

Abstract

Diffusion-based policies have achieved remarkable results in robotic manipulation but often struggle to adapt rapidly in dynamic scenarios, leading to delayed responses or task failures. We present DCDP, a Dynamic Closed-Loop Diffusion Policy framework that integrates chunk-based action generation with real-time correction. DCDP integrates a self- supervised dynamic feature encoder, cross-attention fusion, and an asymmetric action encoder-decoder to inject environmental dynamics before action execution, achieving real-time closed- loop action correction and enhancing the system’s adaptability in dynamic scenarios. In dynamic PushT simulations, DCDP improves adaptability by 19% without retraining while requir- ing only 5% additional computation. Its modular design enables plug-and-play integration, achieving both temporal coherence and real-time responsiveness in dynamic robotic scenarios, including real-world manipulation tasks. The project page is at: https://github.com/wupengyuan/dcdp

Index terms

Manipulation Planning