Since naturalistic behaviors are, in nature, to be long horizon and highly none-linear, none-gaussian with respect to their goals, this sequence modeling problem would be particularly challenging as we want to understand as well generate long biomechanically realistic sequences from modeling the progression of these goal sequences that has only partial observables (keypoints). To this end, advised by professor Scott Linderman at Stanford University, I am doing sequence modeling using MIMIC-MJX, which provide us with observables that was not available previously, to develop deep state space models that can both help us understand the dynamics of behaviors as well to generate them. TopoVNL was presented at the Stanford Undergraduate Research Program Symposium (SURP) 2025 with this poster.
MimicDyn Repository (Closed)