Velodyne LiDAR-Guided Safe Reinforcement Learning for Autonomous Navigation Quality

A goal of safe reinforcement learning is to minimize reward while respecting safety constraints, but achieving that balance remains a challenge. Popular approaches often lead to overly cautious policies that produce suboptimal results in different environments. This paper explores the potential of dual-delay deterministic policy gradients and proximal policy optimization combined with Velodyne VLP-16 LIDAR data to process the data in specific autonomous environment. The surroundings are searched for points of interest (POI) for potential navigation directions, and an ideal waypoint is chosen depending on the data at hand. We deliver comprehensions into improving the performance of Safe RL by enhance perception of receiving data with LiDAR, allowing us to analyze its ability to navigate in dynamic environments while maintaining safety and efficiency using safe reinforcement learning algorithms