Chance Constrained Deep Reinforcement Learning for Autonomous Vehicle Behavioral Planning

Abstract

Autonomous driving is expected to have a major impact on the automotive industry, hence there is a lot of research being done in this area. Autonomous vehicles must be equipped with an efficient and safe planning module that extracts safe policies at all circumstances. Collision avoidance has been one of the complicated tasks in autonomous deriving. Most traditional methods either use optimization techniques or model-based solvers, where it is either not feasible in real-time or requires a full model of the environment to generate policy. Model-free deep reinforcement learning (DRL) however, introduces ways to solve the above-mentioned challenges. This thesis proposes chance-constrained deep reinforcement learning for autonomous vehicle planners which can perform safe driving tasks in a simulated environment. The problem is formulated as chance constrained Markov decision process (CC-MDP) where the constraint is to bound the probability of being in the risky state by some threshold. By incorporating this constraint, the autonomous vehicle is able to learn a safe policy while still achieving its primary objective. This approach effectively balances the trade-off between risk and the successful execution of the objective.

Date of Award	14 Dec 2023
Original language	American English
Supervisor	Majid Khonji (Supervisor)