Learning Generalizable Visuomotor Policy through Dynamics-Alignment

2025-10-30

AI Summary

Behavior cloning suffers from poor generalization due to scarcity of expert demonstrations, while existing video prediction models lack control-input awareness and struggle to support precise manipulation tasks. To address these limitations, this paper proposes Dynamic Alignment Flow Matching (DAP), a novel framework that establishes a mutual feedback loop between a policy network and a dynamics model. DAP employs a dynamics-aware architecture to jointly optimize action generation and state evolution, and introduces self-correcting dynamic alignment during inference to enhance out-of-distribution (OOD) adaptability. Evaluated on real-world robotic manipulation tasks, DAP significantly outperforms mainstream baselines—particularly under OOD conditions such as visual occlusions and lighting variations—demonstrating superior robustness and generalization capability.

Technology Category

• Intelligent Robots: Manipulation
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Computer Vision: Diffusion Models for Vision

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Behavior cloning methods for robot learning suffer from poor generalization due to limited data support beyond expert demonstrations. Recent approaches leveraging video prediction models have shown promising results by learning rich spatiotemporal representations from large-scale datasets. However, these models learn action-agnostic dynamics that cannot distinguish between different control inputs, limiting their utility for precise manipulation tasks and requiring large pretraining datasets. We propose a Dynamics-Aligned Flow Matching Policy (DAP) that integrates dynamics prediction into policy learning. Our method introduces a novel architecture where policy and dynamics models provide mutual corrective feedback during action generation, enabling self-correction and improved generalization. Empirical validation demonstrates generalization performance superior to baseline methods on real-world robotic manipulation tasks, showing particular robustness in OOD scenarios including visual distractions and lighting variations.

Problem

• Learning action-aware dynamics for robotic manipulation tasks
• Improving policy generalization beyond expert demonstration data
• Enhancing robustness against visual distractions and lighting variations

Innovation

• Dynamics prediction integrated into policy learning
• Mutual corrective feedback between policy and dynamics
• Self-correction mechanism for improved generalization