Abstracting Robot Manipulation Skills via Mixture-of-Experts Diffusion Policies

2026-01-29

AI Summary

This work addresses the scalability limitations of existing diffusion-based policies in multitask robotic manipulation, which suffer from large model sizes and high demonstration data costs. To overcome these challenges, the authors propose Skill Mixture-of-Experts Policy (SMP), a novel framework that learns a compact set of orthogonal skill bases and dynamically activates only a few task-relevant experts at each action step, enabling efficient inference and strong generalization. SMP introduces a variational objective–based mixture-of-experts diffusion architecture, augmented with sticky routing and adaptive expert activation mechanisms to significantly reduce computational overhead while enhancing cross-task transferability. Experiments demonstrate that SMP consistently outperforms large diffusion baselines on both simulated and real-world dual-arm platforms, achieving higher task success rates with substantially lower inference costs.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Intelligent Robots: Manipulation
• Planning, Routing, and Scheduling: Learning for Planning and Scheduling

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Diffusion-based policies have recently shown strong results in robot manipulation, but their extension to multi-task scenarios is hindered by the high cost of scaling model size and demonstrations. We introduce Skill Mixture-of-Experts Policy (SMP), a diffusion-based mixture-of-experts policy that learns a compact orthogonal skill basis and uses sticky routing to compose actions from a small, task-relevant subset of experts at each step. A variational training objective supports this design, and adaptive expert activation at inference yields fast sampling without oversized backbones. We validate SMP in simulation and on a real dual-arm platform with multi-task learning and transfer learning tasks, where SMP achieves higher success rates and markedly lower inference cost than large diffusion baselines. These results indicate a practical path toward scalable, transferable multi-task manipulation: learn reusable skills once, activate only what is needed, and adapt quickly when tasks change.

Problem

• robot manipulation
• multi-task learning
• diffusion policies
• scalability
• transfer learning

Innovation

• Mixture-of-Experts
• Diffusion Policy
• Skill Abstraction
• Multi-task Robot Manipulation
• Efficient Inference