Safe Autonomous Navigation Using LiDAR and Camera Fusion with Reinforcement Learning

The primary goal of safe reinforcement learning is to balance reward maximization with adherence to safety constraints, but finding this balance remains a challenging task. Many existing approaches often result in overly cautious policies that lead to suboptimal performance across various environments. In this paper, we propose a technique for autonomous navigation that leverages LiDAR-camera fusion data. By integrating the complementary information from LiDAR and camera sensors, we enhance environmental perception, enabling more accurate and reliable navigation in complex, dynamic settings. Using this fused data, we apply advanced reinforcement learning algorithms specifically Soft Actor-Critic (SAC) and Twin Delayed Deep Deterministic Policy Gradient (TD3) to train autonomous agents for navigation. The synergy between these cutting-edge algorithms and the rich sensor data improves decision-making efficiency, promotes better exploration, and accelerates the learning process in continuous control tasks. Our research underscores the potential of combining LiDAR-camera fusion with reinforcement learning to significantly enhance autonomous navigation system.