Michele Rocca
I am an Industrial PhD Student at the University of Copenhagen (KU). I am finalizing my PhD in Computer Science advised by Professor Kenny Erleben, Professor Sune Darkner, and Professor Sheldon Andrews from l'École de technologie supérieure (ÉTS). I have a background in Biomedical Engineering and Applied Mathematics. I research Motion RL Policies as an inner representation of motion that can both encode its characteristics and be run as a physics simulation in the time domain. My works use the space of the weights of motion policies as a continuous mathematical object approximated from data.
Policy-Space Diffusion for Physics-Based Character Animation
We present a new perspective on physics-based character animation. Assuming policies for similar motions should have similar weights, we introduce regularization during RL training to preserve weight similarity. By modeling the weights’ manifold with a diffusion model, we generate a continuum of policies adapting to novel character morphologies and tasks.
Policy-space Interpolation for Physics-based Characters
We propose an alternative to action interpolation that evaluates only a single policy network per character, regardless of how many trained policies are combined. A graph-based regularization ensures smooth transitions in parameter space, enabling intermediate policies via linear parameter interpolation. This preserves visual quality and enables novel motion variations through weight perturbations.