Reproduce “LSTP-Nav: Lightweight Spatiotemporal Policy for Map-free Multi-agent Navigation with LiDAR” 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
Xingrong Diao, Zhirui Sun, Jianwei Peng, Jiankun Wang
A lightweight spatiotemporal policy, LSTP-Nav, enables map-free multi-agent navigation for resource-constrained robots by processing raw LiDAR inputs, achieving robust collision avoidance and efficient zero-shot sim-to-real transfer. It operates at over 40 Hz on CPU-only platforms, demonstrating high success rates in complex dynamic environments.
Key insights
- A lightweight network combining GRU for temporal encoding and attention for immediate threat focus can effectively process raw LiDAR for real-time, map-free navigation on CPU-only platforms.
- The novel Heading Stability (HS) reward significantly stabilizes training and yields policies that avoid unnecessary detours while maintaining superior collision avoidance in complex multi-agent settings.
- Physics-enhanced simulation paired with localized replay of challenging scenarios effectively boosts learning efficiency and enables robust zero-shot transfer of policies from simulation to physical robots.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.