The passive chicken and aggressive car problem
Published:
The idea was to leverage pedestrian-vehicle interaction at road intersections, to induce a ‘passive-aggressive’ behavior in autonomous vehicles. For this, I employ Inverse Reinforcement Learning (IRL) to explain pedestrain behavior trained on an aggressive car. This learned model explains the pedestrain's motion relative to the car's movement. This project was motivated by Dorsa Sadigh's work on human-aware autonomous cars (see 1, 2, 3).
To get into the details, I first used Deep-Q Networks (DQN) on random pedestrian profiles to learn a car behavior. I then recovered the pedestrian utility by training IRL methods on the learned car model. On learning this pedestrain model, I observed that the pedestrian backs off and waits before proceeding. The car would then accelerate/decelerate accordingly, which is the desired behavior. A short summary of the preliminary results is available here. The project codebase can be found here.
To get into the details, I first used Deep-Q Networks (DQN) on random pedestrian profiles to learn a car behavior. I then recovered the pedestrian utility by training IRL methods on the learned car model. On learning this pedestrain model, I observed that the pedestrian backs off and waits before proceeding. The car would then accelerate/decelerate accordingly, which is the desired behavior. A short summary of the preliminary results is available here. The project codebase can be found here.
