Reproduce “Relational Graph Learning for Crowd 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
Changan Chen, Sha Hu, Payam Nikdel, Greg Mori, Manolis Savva
Relational Graph Learning for Crowd Navigation combines dynamically inferred agent relationships, learned human-motion prediction, and multi-step model-predictive reinforcement learning. In five-person simulated crossing scenarios with non-reciprocal humans, the resulting MP-RGL-Multistep system reaches 0.96 success, lowers collisions to 0.02, and reduces extra travel time to 1.86 seconds, outperforming attention-based, linear-prediction, and one-step planning baselines.
Key insights
- Inferring human-human relationships improves navigation independently of replacing linear motion assumptions: the graph-based linear model reduces collision rate from 0.09 to 0.04 relative to LM-SARL-Linear.
- Learned relational motion prediction helps the robot anticipate coordinated or interacting human behavior that is missed by constant-velocity models.
- Multi-step evaluation resolves situations where an action looks safe immediately but causes a later collision, or where a temporarily close maneuver is preferable to freezing or taking a long detour.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.