Accounting for Unpredictability in Autonomous Driving Behaviour

Abstract

Autonomous Vehicles (AVs) need to behave like humans when interacting with them.We define unpredictability of surrounding drivers as a measure to take into account for trajectory planning and use Maximum Entropy Inverse Reinforcement Learning (IRL) to demonstrate that incorporating unpredictability into a lane change reward function provides insights on human driving behaviour. We first evaluate the IRL algorithm on a Linear Quadratic Regulator proof of concept. Then we use the IRL algorithm to model reward functions for conducting a lane change maneuver in a highway setting. We investigate whether the unpredictability of surrounding traffic will have an effect on the behaviour of the lane changing car by learning two reward functions from human data, a baseline reward function and a reward function that incorporates unpredictability. Our evaluation confirms that incorporating unpredictability results in modest improvements in explaining the behaviour of human drivers and can result in human-like AVs.