Adaptive Path Planning for Autonomous Vehicles in Adverse Driving Conditions and Aggressive Scenarios: A TD3-Based Multimodal Approach

This paper introduces a comprehensive framework that aims to enhance decision making in autonomous vehicles by optimising the Perception and Local Path Planning stages based on the environment surrounding the autonomous vehicle. This framework is designed with the primary purpose of adapting to complex environments such as degraded / pothole-ridden roads, unpredictable weather conditions, and aggressive driver behaviours. It uses a multimodal TD3 Reinforcement Learning algorithm that is fed segmented images obtained from a simulated CARLA environment and outputs a set of control commands that govern how the car behaves on the road. The TD3 reward function incorporates route adherence, speed control, smooth driving, among other enhanced safety measures. Lastly, this paper introduces a novel approach by attempting to predict the actions and positions of traffic vehicles around the ego vehicle and using those predictions to make more informed decisions.