MoE-Loco: Mixture of Experts for Multitask Locomotion

2025-03-11

AI Summary

Legged robots face challenges in generalizing across diverse terrains (e.g., bars, gaps, stairs, slopes, barriers) and locomotion gaits (quadrupedal/bipedal), while multi-task reinforcement learning suffers from gradient interference and poor policy transfer. Method: We propose MoE-Loco—a Mixture-of-Experts (MoE) framework for legged locomotion control—featuring a learnable, task-adaptive routing mechanism that dynamically assigns terrain- or gait-specific subtasks to specialized experts. Experts autonomously specialize during training, mitigating gradient conflict and enabling skill composition and transfer. The method integrates multi-terrain/multi-gait motion modeling, sim-to-real co-training, and end-to-end RL optimization. Contribution/Results: MoE-Loco significantly improves cross-terrain generalization and training efficiency on both simulated and real-world quadrupedal and bipedal robots, achieving robust, adaptive locomotion under a single unified policy—marking the first application of MoE architectures to legged robot control.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Search and Optimization: Learning to Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We present MoE-Loco, a Mixture of Experts (MoE) framework for multitask locomotion for legged robots. Our method enables a single policy to handle diverse terrains, including bars, pits, stairs, slopes, and baffles, while supporting quadrupedal and bipedal gaits. Using MoE, we mitigate the gradient conflicts that typically arise in multitask reinforcement learning, improving both training efficiency and performance. Our experiments demonstrate that different experts naturally specialize in distinct locomotion behaviors, which can be leveraged for task migration and skill composition. We further validate our approach in both simulation and real-world deployment, showcasing its robustness and adaptability.

Problem

• Enables single policy for diverse terrains and gaits
• Mitigates gradient conflicts in multitask reinforcement learning
• Demonstrates expert specialization for task migration and skill composition

Innovation

• Mixture of Experts for multitask locomotion
• Mitigates gradient conflicts in reinforcement learning
• Supports diverse terrains and gaits effectively