Vision-based approaches have dominated the scene of Reinforcement Learning (RL) when applied to autonomous driving agents due to being a cheaper alternative. Addressing a lack of research on other techniques, this project introduces a sensor-based racing environment for multiple agents. The Soft-Actor Critic (SAC) and Proximal Policy Optimization (PPO) RL algorithms are applied to this environment to determine whether on-policy algorithms are better suited towards autonomous sensor-based driving, or if off-policy techniques are superior. Hybrid approaches through Imitation Learning are also explored which apply Generative Adversarial Imitation Learning (GAIL) and Behavioural Cloning (BC). Results indicate that the off-policy SAC models are better suited towards sensor-based racing by adopting a competitive approach. Additionally, these are also shown to be better at avoiding undesired actions such as collisions with other cars. On the other hand, PPO agents have favoured a cooperative approach which led to safer track navigation at the expense of velocity. Such behaviours could all be fully appreciated within the paper's video
Please refer to the project report for further information
Link: https://youtu.be/MEz6-LGSXNk
