Adaptive RL-based AI for real-time dynamic navigation in autonomous robot

Autonomous robot navigation necessitates effective path planning strategies. This project undertakes the implementation and evaluation of a reinforcement learning-based navigation framework using Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC), integrated with multi-sensor fusion and ROS-based control. All algorithms are implemented primarily in Python. A*, Dijkstra, RRT, RRT*, and DQN are evaluated within a custom 2D grid-based simulation environment. Additionally, Dijkstra's algorithm is demonstrated in a 3D physics-based simulation using the PyBullet library. Key performance metrics, including path length, computation time (search/inference), training time and convergence (for DQN), and success rate, are systematically measured and analyzed across various environmental configurations. This work provides a practical comparison of these distinct approaches, highlighting the inherent trade-offs between optimality guarantees, computational efficiency, exploration capabilities, and adaptive learning potential for foundational robotic path planning problems in static environments In simulation, PPO achieved a 92.1% success rate with a mean path length of 14.3 m, while SAC achieved 85.2% and 15.6 m respectively, demonstrating the benefits of the proposed integration.