Reproduce “Nocturne: a scalable driving benchmark for bringing multi-agent learning one step closer to the real world” on OpenResearch

About this paper

Eugene Vinitsky, Nathan Lichtlé, Xiaomeng Yang, Brandon Amos, Jakob Foerster

Nocturne introduces a 2D driving simulator and benchmark that achieves over 2000 steps per second (SPS) by using a C++ backend for efficient partial observability and integrates real-world data from the Waymo Open Motion Dataset. This platform reveals that current multi-agent reinforcement learning methods achieve a 71.7% goal rate but suffer a 20.3% collision rate, especially in highly interactive scenarios, underscoring significant challenges in multi-agent coordination.

Key insights

• Achieving human-level performance in multi-agent driving under partial observability remains a significant challenge, with current methods struggling to generalize across diverse real-world scenarios.
• Policy failures, particularly collisions and inability to reach goals, are most prevalent in highly interactive scenes, indicating that current AI struggles with complex coordination, waiting, and merging behaviors.
• Efficiently modeling human-like partial observability via vectorized observations, rather than computationally expensive camera inputs, enables scalable research into inference and 'theory of mind' in MARL.

Summary by alphaXiv.