Research map / Probabilistic and scientific modeling / Dynamical systems

Support-Safe Variational Hybrid Filtering for Contact-Mode and Sparse-Law Recovery

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: State-space inference for stochastic systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VHyDRO, a support-safe variational hybrid state-space model that mixes the learned proposal with a model-feasible transition carrier, jointly infers continuous latent and discrete contact modes, and fits sparse port-Hamiltonian laws per mode.

Problem

Amortized variational filters for partially observed contact-rich robot dynamics can lose feasible contact branches (zero proposal mass), causing unstable importance weights, mode collapse, and poor interpretable physics recovery.

Approach

Sample from a defensive mixture of the learned proposal and the model's feasible one-step transition law with a pre-sampling mixture weight chosen from a variance certificate, use the same mixture in the importance weight, and combine with a discrete contact mode and mode-conditioned sparse port-Hamiltonian regression.

Key results

• At 90% occlusion on Contact-style tasks, the conservative support-safe variant raises ESS/N from 0.042 (lambda=0) to 0.141 and cuts NLL from 11.286 to 0.161.
• On ManiSkill proxy-label segmentation, VHyDRO reaches ARI 0.647±0.019 and change-point F1 0.652±0.019 versus 0.490 and 0.527 for HMM.
• On four known-equation hybrid systems, VHyDRO-full reaches support F1 0.978±0.042 and relative coefficient error 0.006 versus 0.713 support F1 for no-mode.

Surprising findings

• The adaptive support-safe variant under-covers at low occlusion (Cov90 0.162-0.222), so the gain is not uniform calibration dominance.
• The HMM achieves higher Mode F1 than VHyDRO on PushCube-v1 (0.828 vs 0.810).

Limitations

• Guarantees are model-relative: a physical branch omitted by the carrier cannot be recovered.
• Mode set and sparse library are fixed; action-cancellation assumes logged interventions or mode-independent policies.
• Evaluation uses kinematic proxy labels rather than contact-force ground truth, and no closed-loop task success is reported.

Experimental setup

Datasets: ManiSkill 2 HDF5 demonstrations (PushCube-v1, PokeCube-v1, PullCubeTool-v1, PegInsertionSide-v1) Contact-style tasks (PickCube, PokeCube, PullCube, PushCube) BridgeData-style diagnostics

Benchmarks: ManiSkill 2 demonstrations Sawyer/BridgeData-style diagnostics known-equation hybrid systems

Metrics: ESS/N relative weight variance empirical relative variance NLL ECE Cov90 Mode F1 ARI change-point F1 segment purity

Baselines: VHyDRO no-support (lambda=0) smooth latent HMM no-mode no-sparsity no-pH

Builds on: defensive mixture importance sampling FIVO Lasso port-Hamiltonian systems SINDy

Closest papers by meaning

• Contractive Dynamical Imitation Policies for Efficient Out-of-Sample Recovery (ICLR 2025 · similarity 0.85)
• Geometric Contact Flows: Contactomorphisms for Dynamics and Control (ICML 2025 · similarity 0.85)
• ForceFlow: Learning to Feel and Act via Contact-Driven Flow Matching (NeurIPS 2026 · similarity 0.85)
• Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dyanmics Manifolds (NeurIPS 2026 · similarity 0.85)
• One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.84)
• Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.84)