VHYDRO: Variational Hybrid Filter for Contact-Rich Robot Dynamics

There's a new method called VHYDRO (Variational Hybrid Dynamics Learner) that addresses challenges in robot dynamics, especially when robots interact with various contacts like free movement, impacts, and stick-slip. Regular filters sometimes mess up when they overlook valid contact transitions by assigning them zero probability. VHYDRO improves this by combining learned proposals with a suitable transition law before sampling, ensuring all possible transitions are taken into account. It identifies both a continuous latent state and a discrete contact mode while fitting a sparse port-Hamiltonian law for each situation. This approach guarantees better filtering stability, focuses on coherent contact regimes, and enables simpler port-Hamiltonian representations. You can check out the research on arXiv with the ID 2605.16398.