Reproduce “CE-Nav: Flow-Guided Reinforcement Refinement for Cross-Embodiment Local Navigation” 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

Kai Yang, Tianlin Zhang, Zhengbo Wang, Zedong Chu, Xiaolong Wu, Yang Cai, Mu Xu

AMAP, Alibaba Group's CE-Nav framework enables robust local navigation policies transferable across diverse robot embodiments by combining an offline-learned multi-modal geometric expert with an online reinforcement learning refiner. This dual-stage approach achieves high navigation performance on five distinct robots, including bipedal and quadrupedal platforms, while reducing training time by 8x compared to state-of-the-art end-to-end RL methods and demonstrating strong real-world transferability.

Key insights

  • Decoupling universal geometric reasoning from embodiment-specific dynamic adaptation significantly improves policy transferability and reduces adaptation costs for new robots.
  • Leveraging a conditional normalizing flow model (VelFlow) to learn multi-modal action distributions as an expert prior effectively overcomes the 'disastrous averaging' problem in navigation tasks.
  • A curriculum-based guidance strategy during reinforcement learning allows the agent to efficiently adopt expert knowledge while still enabling sufficient exploration to discover robot-specific optimal behaviors and compensate for low-level controller imperfections.

Summary by alphaXiv.

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