Data-driven simulator of multi-animal behavior with unknown dynamics via offline and online reinforcement learning

This dataset contains the original, unprocessed tracking outputs and accompanying metadata used in our iScience manuscript on AnimaRL, a data-driven simulator of multi-animal behavior. It includes time-stamped trajectories for four domains: synthetic pursuit agents (500 episodes; 10 Hz), flies (107 episodes; 10 Hz), Japanese fire-bellied newts (280 episodes; 10 Hz), and silkmoths (60 episodes; 2 Hz) with per-episode boundaries and experimental-condition labels where applicable (two-condition setups for agents and silkmoths). For each episode, the files provide the raw recorded variables (e.g., positions/velocities for agents and insects; and for silkmoths, body angle and sensory measurements such as odor/wind/vision alongside kinematics) . These raw files are intended to support transparent reuse and independent re-analysis; scripts for preprocessing, model training, and reproducing the analyses reported in the manuscript are available in the companion code repository (https://github.com/keisuke198619/animarl).