Learning Realistic and Safe Pedestrian Behavior by Imitation

Navigating amongst pedestrians is a very complex task for an autonomous agent. Not only must the agent understand traffic rules and navigate safely, but it must also act in a way that obeys social norms and does not interfere with other pedestrians. Here, we focus on obtaining a model that portrays pedestrian behavior from real-life demonstrations using imitation learning. We create a reinforcement learning environment that allows an agent to learn to navigate using the obtained behavior model. The work is still in progress, but we illustrate how we generate demonstrations of pedestrian behavior from video captured by smart glasses and we incorporate them into a reinforcement learning environment.