Reproduce “REASAN: Learning Reactive Safe Navigation for Legged Robots” 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

Qihao Yuan, Ziyu Cao, Ming Cao, Kailai Li

REASAN introduces a modularized end-to-end framework for LiDAR-based reactive safe navigation, enabling legged robots to perform real-time, collision-free movement in complex dynamic environments. The system, demonstrated on a Unitree Go2 robot, achieves 50 Hz onboard performance, successfully navigating diverse real-world scenarios including multi-robot interactions and outperforming monolithic approaches in simulation with 95.2% success in complex maze tasks.

Key insights

  • Modular decomposition of the complex legged motor-control problem simplifies learning, allowing for lightweight neural networks and standard RL algorithms, which enhances robustness and tractability.
  • Isolating a 'safety-shield' policy to reactively modify velocity commands based on immediate perception effectively handles dynamic obstacles, without needing complex policy-switching or intermediate heuristics.
  • Using privileged information, such as an occupancy grid-based guiding direction, during training is effective for addressing sample inefficiency and sparse rewards in complex goal-reaching navigation tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.