In a study published in Robot Learning journal, researchers propose a new learning-based path planning framework that allows mobile robots to navigate safely and efficiently using a Transformer model. By learning from Improved RRT* with Reduced Random Map Size path-planning algorithms and combining this knowledge with a modified right-of-way rule, the system enables reliable navigation and replanning in dynamic multi-robot environments.
Autonomous mobile robots are increasingly used in factories, warehouses, and service environments to transport materials and perform repetitive tasks. To operate safely, robots must plan collision-free paths while adapting to unexpected obstacles and interactions with other robots. Traditionally, this requires complex navigation pipelines that combine mapping, localization, and planning, often demanding high computational resources.
In this study, researchers introduce a learning-based alternative called the Path Planning Transformer (PPT). Instead of relying on continuous online mapping, the model learns how to plan efficient paths directly from occupancy maps by observing expert trajectories generated using an improved version of the classical RRT* algorithm.
"Classical planners such as A* or RRT* are reliable but can struggle to re-plan smoothly in dynamic environments, especially when multiple robots are involved," explains the author. "Our goal was to teach a model to learn these planning behaviors and reproduce them efficiently in real time."
The PPT model was trained on thousands of automatically generated path examples, each demonstrating how an optimal trajectory avoids obstacles. Once trained, the model predicts smooth paths using a Transformer architecture, a type of neural network originally developed for language processing and now increasingly applied in robotics.
To handle interactions between robots, the system introduces a modified right-of-way rule. When a robot detects another robot or an unexpected obstacle using its LiDAR sensor, the map is updated with a virtual obstacle that enforces a preferred passing direction. This simple mechanism allows each robot to re-plan independently, without explicit communication or centralized coordination.
The approach was evaluated in both simulation and real-world experiments using two mobile robots. Results show that the learning-based planner consistently produced smoother paths with fewer turns than traditional A* and RRT* methods, particularly after replanning. While classical planners sometimes generated shorter paths, these often involved sharp turns or complex maneuvers that are less suitable for real robots.
"All experiments and simulations were carried out on a standard laptop using MATLAB, ROS, and Gazebo," the author adds. "This demonstrates that the system is practical and does not require specialized hardware."
The findings suggest that learning-based planners can complement classical algorithms by improving path smoothness and adaptability while keeping computational requirements low. This has potential applications in industrial automation, warehouse robotics, and collaborative robot systems.
While the current study focuses on two-robot scenarios in 2D environments, future work will explore extending the approach to larger robot teams and three-dimensional navigation using voxel-based maps.
This paper was published in Robot Learning. Lonklang A, Botzheim J. Path Planning Transformers supervised by IRRT*-RRMS for multi-mobile robots. Robot Learn. 2026(1):0005, https://doi.org/10.55092/rl20260005.
DOI: 10.55092/rl20260005
