This paper proposes a learning-based path planning framework for autonomous mobile robotsbased on a novel Transformer architecture supervised by an Improved RRT* with Reduced Random MapSize (IRRT*-RRMS) algorithm. The system, referred to as the Path Planning Transformer (PPT), predictssequences of intermediate waypoints from occupancy maps and eliminates the need for traditional graph-basedalgorithms or SLAM during online execution. The model is trained using a dataset of 10,000 samples generatedfrom 100 randomly created maps, each containing 100 distinct start-goal trajectories. Ground truth pathsare computed using IRRT*-RRMS, which enhances classical RRT* through dynamic sampling regionreduction, node deletion, and bacterial mutation-based smoothing.

    To ensure robustness in dynamic environments, a modified right-of-way rule is introduced as a post-planning collision avoidance mechanism. This rule employs a 270-degree virtual obstacle ring projected in front of the robot upon LiDAR-based detection of unknown dynamic obstacles. It enables the system to trigger safe re-planning that biases the trajectory away from oncoming objects—especially in multi-robot scenarios. Real-world experiments using two TurtleBot3 Waffle Pi robots demonstrate the framework’s capability to dynamically avoid collisions while maintaining progress toward assigned goals. In scenarios where robots initially face each other, the system successfully replans their paths without direct communication, relying solely on onboard sensing and map-based planning.

    The PPT model was trained for 429 epochs, achieving a best validation loss of 186. Experimental results confirm that the approach generates efficient, smooth, and collision-free trajectories in both simulation and physical deployment. The proposed framework provides a scalable solution for map-driven navigation in indoor environments and opens the door to future integration with higher-level task-planning and semantic-reasoning systems.