Addressing Reward Hacking in Autonomous Navigation using a Soft Actor Critic approach

In order to address the challenges of the dynamic environment in urban mobility, Reinforcement Learning (RL) has gained a lot of attention as a viable solution, due to its ability to handle the complexity of the related task. In this context, the scope of the proposed work is twofold. Initially, by adopting the RL-based approach, an autonomous agent is developed, that is able to navigate an electric vehicle in an urban environment using the Soft Actor Critic (SAC) algorithm. Subsequently, reward hacking and its impact on the overall behavior of the autonomous vehicle is investigated, since it may lead to undesirable behaviors of the trained agent. The proposed approach is demonstrated using a specialized simulated framework, along with a custom gymnasium environment for the training of the autonomous agents.