End-to-end Multi-Objective Deep Reinforcement Learning for Autonomous Navigation

Autonomous navigation with an end-to-end reinforcement learning paradigm for ground vehicles poses significant safety and multi-objective challenges, limiting its practical implementation in real-world scenarios. This research proposes a reinforcement learning approach to address the challenges of end-to-end navigation policy that accounts for dynamic multi-objectives along with safety concerns. An action selection law is designed to ensure smooth sequential actions. The dynamic weighting of multiple objectives enhances adaptivity in policy learning. To improve safety, intensive rewards are used to penalize sparse risky actions. The proposed approach is validated using three deep reinforcement learning frameworks in a 2D world navigation task of pursuing dynamic goals while avoiding obstacles. This research presents a promising solution to achieve a multi-objective end-to-end policy for handling dynamic and complex scenarios.