A Large-Scale Dataset for Training Microrobot Navigation Policies

This dataset contains a large-scale collection of artificial vascular environments developed for training and evaluating microrobot navigation policies. The dataset was derived from the dataset accompanying Autonomous Environment-Adaptive Microrobot Swarm Navigation Enabled by Deep Learning-Based Real-Time Distribution Planning (https://doi.org/10.6084/m9.figshare.19149779.v1). The original dataset contains 43,133 samples, each comprising an artificial vascular channel and a simulated microswarm. The vascular channels were generated to capture representative characteristics of real vascular environments, including variations in channel diameter, branched structures, and open-sided channels with curved boundaries. To construct the present dataset, the simulated microswarms were removed from the original samples, leaving isolated vascular environments. Duplicate environments were subsequently identified and removed, resulting in 14,753 unique vascular channel instances. During policy training, the initial and target positions of the microrobot are randomly sampled within the navigable vascular space of each environment. This design enables the dataset to support large-scale training and systematic evaluation of microrobot navigation methods across diverse vascular geometries.