Tree-LSTM and Transformer-Enhanced RRT for Efficient Path Planning

The rapidly exploring random tree star (RRT*) algorithm is widely used for path planning, but finding an initial feasible path can require substantial sampling. Learning-based enhancements commonly predict promising sampling regions from grid maps. This paper presents a model that combines a Tree Long Short-Term Memory encoder with a Transformer decoder to predict an ordered sequence of keypoints from a hierarchical map representation. The resulting TLT-RRT* planner checks the predicted segments, repairs infeasible connections, and continues path optimization using RRT*. The multi-branch tree representation avoids explicitly storing a full high-dimensional grid when the environment admits a compact hierarchical representation. Two- and three-dimensional experiments show shorter initial-path times than the selected RRT-based baselines in the evaluated settings. Six-dimensional manipulator examples and resource benchmarks further characterize the implementation and its computational limits.