Understanding how animals transition between behavioral modes has been a central problem in neuroethology. However, naturalistic data makes it hard: transitions vary at different scales and conventionally segmentation algorithms lack a control-relevant grounding. Recent deep reinforcement learning imitation policies emulate animals in a biomechanical simulator, exposing a learned latent intention space and a rolled-out joint-angle (qpos) space that are both egocentric and more biomechanically realistic than the underlying keypoints. This latent space provides a meaningful feature representation for the behavior data, which we aim to analyze behavior movement trajectories in.
To this end, we first map frame-by-frame trajectories in this latent space to density-weighted point clouds. We then apply a topological method, called the discrete-Morse based graph construction method (abbrev. DM-graph) to extract two meaningful graph skeleton primitives from the density point clouds: DM-paths, which intuitively are the low-density corridors the trajectory follows between dense behavioral regions, decomposed by density persistence so that a long corridor stays intact while subtler changes peel off as their own shorter branches, and DM-cycles, which intuitively captures behaviors whose entry and exit arms occupy spatially distinct routes. These DM-skeletons provide a better segmentation of behaviors in individual trajectories. We further cluster these behaviors by clustering these DM-skeletons based on the Wasserstein distance. Empirically, on a rat behavior dataset across three feature spaces (egocentric keypoints, joint angles, and a learned latent), Wasserstein clustering of the primitives exposes transition structure that prior state space models, deep autoencoders, and topological clustering (Mapper) baselines do not produce. Several transitions appear consistently across all three feature spaces, indicating the subspace itself is a topological invariant of the trajectory.
I am working on this with advising from professor Yusu Wang and professor Talmo Pereira. This work is currently under submission. TopoMIMIC was presented at the Annual Society for Neuroscience (SfN) 2025 Conference with this poster.