← Back ICRA 2026

MetaDP: Meta-Manipulation Diffusion Policy for Robotic Manipulation

Zheyi Zhao,, Ying He,, F. Richard Yu, Jiyuan Song, and Xilong Xun

PDF

AI summary

Key figure (auto-extracted from paper)

Integrating discrete task-phase vectors into a diffusion policy significantly reduces compounding errors and boosts precision in robotic manipulation.

Robotic manipulation Diffusion policy Compounding errors Task phase awareness Imitation learning High-precision control

Problem

Existing end-to-end imitation learning methods for 3D robotic manipulation lack explicit task-phase awareness, causing compounding errors and poor performance in high-precision tasks where minor deviations lead to failure.

Approach

MetaDP injects trainable meta-manipulation prompt vectors indicating the current task stage into a diffusion policy, dynamically modulating the denoising process alongside noise levels to guide precise action prediction.

Key results

5% relative improvement over baselines across eight complex tasks
Mitigation of compounding errors via phase-consistent denoising control
Clear separation of learned task-phase representations in latent space
Superior performance in high-precision simulation benchmarks

Why it matters

Enables more reliable and precise robotic manipulation for high-accuracy automation tasks, advancing the practical deployment of diffusion-based imitation learning.

Abstract

In the field of 3D object manipulation, collecting expert data for end-to-end imitation learning is a standard approach. While successful, previous works have not adequately determined the specific phase of the current task, leading to low tolerance for feature variability and difficulty in addressing compounding errors, which are issues particularly prominent in high-precision tasks. To address these limitations, we introduce a novel framework named Meta-manipulation Diffusion Policy (MetaDP). This framework utilizes meta-manipulation prompt vectors to inform the model of the current stage of the task, which, in conjunction with noise reduction levels, controls the denoising process of the diffusion policy, thereby enhancing the model’s predictive accuracy. Comprehensive experiments demonstrate that MetaDP significantly surpasses established baselines, achieving a relative improvement of 5% across eight complex tasks, which underscores the superiority of our method.

Index terms

Imitation Learning Manipulation Planning AI-Based Methods