MoRE: Mixture of Residual Experts for Humanoid Lifelike Gaits Learning on Complex Terrains

Dewei Wang1,2, Xinmiao Wang2,3, Xinzhe Liu2,4, Jiyuan Shi2, Yingnan Zhao3
Chenjia Bai2† Xuelong Li2
1University of Science and Technology of China 2Institute of Artificial Intelligence (TeleAI), China Telecom 3Harbin Engineering University 4ShanghaiTech University
†Corresponding Author
图片3 图片1 图片4 图片2
arXiv Code

Video

Abstract

Humanoid robots have demonstrated robust locomotion capabilities using Reinforcement Learning (RL)-based approaches. Further, to obtain human-like behaviors, existing methods integrate human motion-tacking or motion prior in the RL framework. However, these methods are limited in flat terrains with proprioception only, restricting their abilities to traverse challenging terrains with human-like gaits. In this work, we propose a novel framework using mixture of latent residual experts with multi-discriminators to train a RL policy, which is capable of traversing complex terrains in controllable lifelike gaits with exteroception. Our training pipeline, containing two stages, first teaches the policy of traversing complex terrains using a depth camera, then enables gait-commanded switching between human-like locomotion patterns. We also design gait rewards to adjust human-like behaviors like robot base height. Simulation and real-world experimental results demonstrate that our framework exhibits exceptional performance in traversing complex terrains, and achieves seamless transitions between multiple human-like gait patterns.

Method

图片1

Experiments


Single Terrain

Plain | Walk & Run gait
Plain | High knees gait
Plain | Squat gait
Pit | Walk & Run gait
Pit | High knees gait
Pit | Squat gait
Stair | Walk & Run gait
Stair | High knees gait
Stair | Squat gait
Gap | Walk & Run gait
Stair | High knees gait
Stair | Squat gait

Combined Terrain


BibTeX


      @misc{wang2025moremixtureresidualexperts,
        title={MoRE: Mixture of Residual Experts for Humanoid Lifelike Gaits Learning on Complex Terrains}, 
        author={Dewei Wang and Xinmiao Wang and Xinzhe Liu and Jiyuan Shi and Yingnan Zhao and Chenjia Bai and Xuelong Li},
        year={2025},
        eprint={2506.08840},
        archivePrefix={arXiv},
        primaryClass={cs.RO},
        url={https://arxiv.org/abs/2506.08840}, 
  }