DTSP Dataset for KAJAN

This dataset provides the fixed benchmark test instances used for evaluating the Kinematics-Aware Joint Attention Network (KAJAN) on the Dubins Traveling Salesman Problem (DTSP). The dataset is released to support reproducibility of the experiments reported in the associated study and to facilitate comparison with other neural and optimization-based routing methods. The dataset contains six fixed test sets, covering three problem sizes (10, 30, and 50 target points) and two spatial distributions (uniform and clustered). Each test set contains 1,000 DTSP instances: - test_uniform_10.pt: 1,000 uniformly distributed instances with 10 targets. - test_uniform_30.pt: 1,000 uniformly distributed instances with 30 targets. - test_uniform_50.pt: 1,000 uniformly distributed instances with 50 targets. - test_cluster_10.pt: 1,000 clustered instances with 10 targets. - test_cluster_30.pt: 1,000 clustered instances with 30 targets. - test_cluster_50.pt: 1,000 clustered instances with 50 targets. Each instance contains the two-dimensional coordinates of the target points and a precomputed Dubins transition-cost matrix. The transition matrix represents the motion cost between target-heading configurations under a fixed minimum turning radius of 0.1 and 16 discrete heading candidates. The datasets are stored in PyTorch (.pt) format and can be loaded directly using torch.load() in a compatible Python and PyTorch environment. Uniform instances are generated by sampling target coordinates within the predefined planar workspace. Clustered instances are generated around multiple randomly positioned cluster centers with bounded Gaussian perturbations, producing spatially concentrated target distributions. Fixed random seeds are used for the released test sets so that the benchmark instances are deterministic and reproducible. The six test sets are used in the associated KAJAN study for evaluating solution quality, inference efficiency, and generalization across problem sizes. The model is trained on 30-target instances and evaluated without retraining on 10- and 50-target instances. The released data therefore cover both the training scale and unseen problem sizes used in the experimental evaluation. This dataset is intended for research in neural combinatorial optimization, the Dubins Traveling Salesman Problem, non-holonomic vehicle routing, motion planning, and physics-aware learning. The files may be used for benchmarking, reproducibility studies, algorithm comparison, and further research on learning-based solutions for kinematically constrained routing problems.