Reproduce “Learning Perceptive Humanoid Locomotion over Challenging Terrain” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wandong Sun, Baoshi Cao, Long Chen, Yongbo Su, Yang Liu, Zongwu Xie, Hong Liu
Researchers from Harbin Institute of Technology developed the Humanoid Perception Controller (HPC), an architecture integrating terrain perception with sensor denoising through a learned world model. This framework enabled a humanoid robot to autonomously traverse 2 kilometers of diverse outdoor terrains, including stairs and rough surfaces, while demonstrating superior velocity tracking and maintaining stability even under extreme sensor noise conditions.
Key insights
- A learned world model with a variational information bottleneck effectively filters noise from exteroceptive perception and robustly estimates the robot's state, crucial for adapting to terrain variability.
- Teacher-student distillation provides an efficient paradigm for transferring optimal behaviors from an oracle with privileged information to a student policy that operates on noisy, real-world-like observations.
- Realistic modeling of diverse sensor noise and comprehensive domain randomization are essential for training humanoid policies that generalize effectively from simulation to real-world deployment without fine-tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.