Reproduce “Advancing Humanoid Locomotion: Mastering Challenging Terrains with Denoising World Model Learning” on OpenResearch

About this paper

Xinyang Gu, Yen-Jen Wang, Xiang Zhu, Chengming Shi, Yanjiang Guo, Yichen Liu, Jianyu Chen

This paper presents the Denoising World Model Learning (DWL) framework, which enables XBot humanoid robots to achieve robust, zero-shot sim-to-real locomotion on challenging terrains using only proprioceptive sensors. The system integrates advanced active ankle control with a deep learning approach that filters sensor noise and estimates terrain properties, demonstrating successful navigation on snowy ground, stairs, and irregular surfaces.

Key insights

• Denoising World Model Learning (DWL) effectively compensates for noisy sensor data and IMU drift, allowing robust state and terrain estimation from proprioceptive sensors alone.
• Combining an active 2-DoF Closed Kinematic Chain Ankle Mechanism with a learning-based controller significantly enhances stability and adaptability on irregular surfaces and during disturbances.
• Extensive domain randomization and noise injection during simulation training enable zero-shot transfer of complex locomotion policies to physical humanoid robots without fine-tuning.

Summary by alphaXiv.