Multi-Agent RL for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty | ICRA 2026

Multi-Agent RL for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty | ICRA 2026
DESCRIPTION
===========
Supplementary video for the paper "Multi-Agent Reinforcement Learning for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty," accepted and presented at the 8th Workshop on Long-term Human Motion Prediction at ICRA 2026.

We co-train a self-driving car (SDC) and 12 pedestrians with Multi-Agent Proximal Policy Optimization (MAPPO). Pedestrian locomotion follows scripted shortest-path navigation, while a reinforcement learning policy controls the high-level go/wait crossing decision. Whether a pedestrian crosses at a crosswalk or jaywalks across the road is driven by a latent personality trait that is hidden from the vehicle, making jaywalking a tunable source of behavioral uncertainty.

Key results (500-episode evaluations):
- The co-trained SDC reaches 78% goal success at a 14% collision rate, versus 35%/33% for the best non-learning baseline and 65%/20% for a single-agent SDC.
- Co-training reduces collisions by about 30% relative to single-agent training, as pedestrians learn to wait when the SDC approaches at speed.
- A speed-differential metric shows the SDC travels 2.65 m/s faster near jaywalkers than near crosswalk users at close range, indicating that jaywalking encounters are harder to anticipate.
- Jaywalking accounts for only 13% of crossings but 62% of collisions.

What you see in the video: the blue rectangle is the SDC; coloured dots are pedestrians (hue encodes jaywalking tendency, green for cautious to red for reckless; yellow segments mark active jaywalking). The clips show both a collision with a jaywalker and successful avoidance maneuvers.

Paper (arXiv): https://arxiv.org/abs/2605.20255
Code: https://github.com/prakash-aryan/marl-sdc-pedestrian-uncertainty

Authors: Prakash Aryan, Kaushik Raghupathruni, Timo Kehrer, Sebastiano Panichella (University of Bern; AI4I, The Italian Institute of Artificial Intelligence).

Citation:
Aryan, P., Raghupathruni, K., Kehrer, T., Panichella, S. (2026). Multi-Agent Reinforcement Learning for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty. arXiv:2605.20255.

#ReinforcementLearning #MultiAgentRL #AutonomousDriving #SelfDrivingCars #MAPPO #ICRA2026 #Robotics #JAX

Видео Multi-Agent RL for Safe Autonomous Driving Under Pedestrian Behavioral Uncertainty | ICRA 2026 канала Prakash Aryan