MARL Enhances Realism in Autonomous Driving Safety Tests with Pedestrian Uncertainty
A recent publication on arXiv (2605.20255) suggests that multi-agent reinforcement learning (MARL) can enhance the simulation-based evaluation of self-driving cars (SDCs) by accounting for the uncertainty in pedestrian behavior. Conventional simulations often utilize scripted or simplified models of pedestrians, which do not accurately reflect the diversity of human actions, particularly in jaywalking situations influenced by underlying personality traits. The study involves the joint training of an SDC alongside 12 pedestrians through Multi-Agent Proximal Policy Optimization (MAPPO). While pedestrian movement adheres to scripted Dijkstra pathfinding, an RL policy governs higher-level decisions related to crossing. The researchers believe that MARL can yield more authentic interactions compared to static pedestrian policies and that the disparity in behavior between predictable and unpredictable crossings can be assessed through trajectories. This research aims to improve the realism of safety evaluations for autonomous vehicles.
Key facts
- arXiv paper 2605.20255 proposes MARL for SDC safety testing
- Jointly trains SDC and 12 pedestrians using MAPPO
- Pedestrians use Dijkstra pathfinding with RL policy for crossing
- Addresses jaywalking scenarios with latent personality traits
- Aims to measure behavior gap between predictable and unpredictable crossings
- Criticizes scripted pedestrian models for lacking heterogeneity
- Focuses on simulation-based testing realism
- Published as a cross-type announcement on arXiv
Entities
Institutions
- arXiv