Multi-Agent Reinforcement Learning for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty

Researchers demonstrate that multi-agent reinforcement learning (MARL) significantly improves autonomous vehicle safety testing by co-training self-driving cars alongside realistic pedestrian agents with hidden behavioral traits. The co-trained SDC achieved 78% goal success with 14% collision rate versus 35%/33% for rule-based baselines, with jaywalking accounting for 62% of collisions despite representing only 13% of crossing events.

This research addresses a critical gap in autonomous vehicle safety validation—the unrealistic nature of scripted pedestrian models in simulation environments. Traditional testing relies on predictable crossing behaviors that fail to capture the heterogeneous decision-making and personality-driven variations observed in real-world pedestrians, particularly concerning unpredictable jaywalking. By implementing MARL with Multi-Agent Proximal Policy Optimization (MAPPO), the team created emergent pedestrian behaviors governed by latent traits invisible to the SDC, more accurately simulating the uncertainty autonomous vehicles face on actual roads.

The experimental results reveal substantial performance improvements in the co-trained system. The 30% collision reduction compared to single-agent RL directly demonstrates that pedestrians learn safer waiting behaviors when exposed to realistic vehicle dynamics, while SDCs simultaneously adapt to genuinely unpredictable crossing patterns. The speed differential metric—showing 2.65 m/s faster travel near jaywalkers—provides quantitative evidence that jaywalking scenarios remain fundamentally challenging for current autonomous systems despite co-training improvements.

Jaywalking's disproportionate collision representation (62% of collisions from 13% of events) signals that autonomous vehicle safety certification requires testing methodologies capturing behavioral unpredictability. This research validates simulation-based multi-agent approaches as superior to traditional scripted models for identifying critical safety gaps. The methodology has immediate applicability to autonomous vehicle development pipelines and regulatory testing frameworks seeking more realistic validation environments that expose edge cases before real-world deployment.

• →Co-trained MARL systems achieve 78% goal success versus 35% for rule-based pedestrian models, demonstrating significant realism improvements in AV testing
• →Jaywalking accounts for 62% of collisions despite being only 13% of crossing events, highlighting vulnerability to unpredictable pedestrian behavior
• →Multi-agent reinforcement learning enables pedestrians to learn realistic wait behaviors, reducing collisions by 30% relative to single-agent approaches
• →Speed differential metrics quantify that SDCs fail to anticipate jaywalking scenarios, traveling 2.65 m/s faster near unpredictable crossers
• →Hidden behavioral traits in pedestrian agents create emergent complexity that better simulates real-world crossing unpredictability than traditional scripted models