← Back
ICRA 2026
MoE-DP: An MoE-Enhanced Diffusion Policy for Robust Long-Horizon Robotic Manipulation with Skill Decomposition and Failure Recovery
Baiye Cheng, Tianhai Liang, Suning Huang, Maanping Shao, Feihong Zhang, Botian Xu, Zhengrong Xue, Huazhe Xu
AI summary
MoE-DP significantly improves robustness in long-horizon robotic manipulation by decomposing policies into specialized, interpretable skills that enable automatic failure recovery and flexible subtask reordering.
Problem
Standard diffusion policies lack stage-awareness and fail to recover from intermediate subtask failures in long-horizon tasks, often causing cascading errors due to entangled, uninterpretable representations.
Approach
The method inserts a Mixture of Experts layer between the visual encoder and diffusion model to dynamically route observations to specialized expert networks, enabling skill decomposition and recovery without re-training.
Key results
- 36% average relative success rate improvement under simulated disturbances
- Significant real-world robustness gains across multiple manipulation tasks
- Clear mapping between specialized experts and semantic task primitives
- Inference-time subtask rearrangement and generalization without re-training
Why it matters
Provides a scalable, interpretable framework for robust long-horizon robotic manipulation, enabling flexible high-level control and reliable recovery for real-world deployment.
Abstract
No abstract on file.