Simulation-Based Validation of a TD3–RRT Hybrid Learning Framework for Safe and Adaptive Robot Navigation

Autonomous navigation in dynamic environments remains a major challenge due to real-time constraints and unpredictable obstacles. Unlike prior work that relies solely on local planning or suffers from unstable convergence, our method aims to ensure safe and efficient navigation for autonomous robots in dynamic environments through a more stable architecture. The research contribution is a TD3+RRT hybrid architecture enhanced with transfer learning, enabling rapid adaptation in dynamic environments while maintaining trajectory feasibility. The TD3 agent is first trained in a static ROS–Gazebo environment and then transferred to a larger, dynamic setting. RRT operates in parallel to generate feasible global paths, while TD3 handles real-time obstacle avoidance. The state space includes LIDAR data and relative goal positioning, and the action space produces continuous velocity commands. The hybrid model achieved a 92% success rate, a 50% reduction in collision rate, and the highest cumulative reward compared to baseline TD3, DDPG, and SAC methods. Trajectory analysis further confirmed smoother and more consistent paths in dynamic scenarios. These results highlight the effectiveness of combining model-based and model-free strategies for reliable autonomous navigation, offering a promising step toward real-world deployment.