VHYDRO: New AI filter prevents robots from losing track in contact tasks

Contact-rich robot dynamics are notoriously tricky—a single sensor reading can map to multiple possible states (free motion, impact, stick-slip). Traditional amortized filters that assign zero probability to a feasible contact transition permanently lose the branch the robot actually follows, causing catastrophic failure. Researchers from the robotics community (authors Marios Papamichalis and Regina Ruane) have introduced VHYDRO (Variational Hybrid Dynamics Learner), a filtering approach that prevents this branch loss. At each step, VHYDRO mixes a learned proposal with a feasible transition law before sampling and importance weighting, ensuring every possible transition remains covered. The system jointly infers a continuous latent state and a discrete contact mode, then fits a sparse port-Hamiltonian law to each regime.

Empirical results demonstrate VHYDRO's robustness. Under heavy occlusion, the support-safe filter stays usable while a non-defensive proposal collapses. On ManiSkill demonstrations and four Sawyer/BridgeData task families, the discrete state forms temporally coherent contact-regime segments that yield stronger joint profiles across Adjusted Rand Index (ARI), change-point F1, and segment purity compared to post-hoc and mode-free baselines. Additionally, on hybrid systems with known equations, the mode-conditioned sparse fit recovers the active physical terms, whereas purely predictive baselines do not. This work provides three theoretical guarantees—support coverage stabilizes filtering, the stabilized filter concentrates the discrete contact posterior, and mode-pure segments admit sparse recovery—all tightly linked to the empirical performance.