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

Problem Statement

Contact-rich robot dynamics are inherently hybrid, characterized by transitions between distinct regimes such as free motion, impact, sticking, and slip. A critical failure mode in standard amortized variational filtering is support mismatch: a learned proposal distribution may assign zero or negligible probability to a feasible contact transition that the robot actually follows. If this branch is deleted early in the filtering process, subsequent observations cannot recover it, leading to catastrophic filtering collapse. This issue is particularly acute under partial observation (e.g., heavy occlusion), where multiple contact-consistent futures remain feasible.

Existing approaches face a trade-off: scalable variational latent models often lack explicit hybrid structure and physical interpretability, while physically structured models or particle-based methods are often computationally expensive to amortize or difficult to deploy as end-to-end learned perception models. Furthermore, standard proposals can fail to cover the target support, resulting in unstable importance weights and degraded learning.

Methodology: VHyDRO

The authors propose VHyDRO (Variational Hybrid Dynamics with Robust Proposal Support), a support-safe hybrid variational state-space model designed for partially observed contact dynamics.

Model Architecture

VHyDRO represents the system state $x_t = (s_t, z_t)$, where:

• $z_t$ is a continuous latent belief state.
• $s_t$ is a discrete predictive contact regime (mode).

The transition dynamics for each mode are parameterized by a port-Hamiltonian core augmented with sparse library corrections. This structure ensures the learned dynamics are physically structured, interpretable, and capable of recovering governing equations.

Support-Safe Inference

The core innovation lies in the inference procedure, which prevents the deletion of model-feasible branches.

1. Defensive Mixture Proposal: Instead of sampling solely from an amortized proposal $Q_t$, VHyDRO samples from a mixture distribution:
   $$Q_{t,\lambda} = (1 - \lambda)Q_t + \lambda P_t$$
   Here, $P_t$ is the model's one-step transition law (the "feasible transition carrier") which defines the set of model-feasible modes and transitions. The learned proposal $Q_t$ is observation-adaptive and sharp, while $P_t$ ensures coverage of the feasible support.
2. Pre-Sampling Budgeting: The mixing weight $\lambda_t$ is chosen before sampling particles, based on a certificate $\bar{\rho}_t$ derived from the desired variance budget. This ensures that the ratio of the target density to the proposal density is bounded ($dP_t/dQ_{t,\lambda} \le 1/\lambda$).
3. Consistent Weighting: The same realized mixture $Q_{t,\lambda}$ is used for both sampling and computing importance weights, maintaining the validity of the Sequential Importance Weighted (IW) / Filtering Variational Objective (FIVO) estimation.

Theoretical Guarantees

The paper establishes three theoretical links connecting the filtering mechanism to downstream capabilities:

1. Support-Safe Filtering (Theorem 1): The defensive mixture ensures that the proposal $Q_{t,\lambda}$ dominates the target support $P_t$. This bounds the relative variance of the normalizer estimate and guarantees a lower bound on the Effective Sample Size (ESS), preventing the collapse of the filter under high occlusion.
2. Mode Concentration (Theorem 2): Under conditions of predictive separation (sufficient log-likelihood gap between modes), the stabilized filter concentrates the discrete posterior on coherent contact regimes. This allows the model to identify temporally consistent segments of contact behavior.
3. Sparse-Law Recovery (Theorem 3): Once mode-pure segments are identified, the paper applies Lasso oracle bounds to show that sparse port-Hamiltonian regression can recover the active physical terms. The error is decomposed into filtering, derivative, mode-impurity, and plug-in perturbation components.

Key Contributions

1. Support-Safe IW/FIVO Gate: A deployable mechanism for amortized hybrid filtering that mixes the learned proposal with a feasible carrier before sampling. Unlike static defensive mixture inequalities, this approach operates sequentially, choosing $\lambda_t$ based on a pre-sampling certificate and using the same mixture for weighting.
2. Predictive Contact Modes: The method jointly infers continuous latent states and discrete contact modes, yielding temporally coherent segments that align with kinematic proxy regimes, outperforming post-hoc segmentation baselines.
3. Controlled Sparse-Law Audit: The framework enables the recovery of interpretable sparse physical laws (port-Hamiltonian coefficients) from mode-conditioned segments, separating accurate prediction from mechanism recovery.

Experimental Results

The authors validate the three mechanism links through three experiments:

Experiment 1: High-Occlusion Support-Mismatch Audit

• Setup: Evaluated on Contact-style tasks (PickCube, PokeCube, etc.) and Sawyer/BridgeData families under varying levels of occlusion (up to 90%).
• Findings: Under heavy occlusion, non-defensive baselines ($\lambda=0$) and smooth latent models suffer from "predictive collapse," characterized by a sharp drop in ESS/N and a massive increase in Negative Log-Likelihood (NLL).
• Performance: The support-safe VHyDRO variants maintain usable ESS/N (e.g., improving from 0.042 to 0.141 at 90% occlusion) and prevent NLL collapse (reducing NLL from ~11.3 to ~0.16). The conservative variant (higher fixed $\lambda$) offers the most robustness, while the adaptive variant balances sharpness and stability.

Experiment 2: Contact-Regime Segmentation

• Setup: Applied to ManiSkill and Sawyer/BridgeData demonstrations. Evaluation uses denoised kinematic proxy labels (free, impact, stick/slip) derived from state/action fields, not simulator contact forces.
• Findings: VHyDRO produces temporally coherent mode segments that align better with proxy labels than baselines.
• Metrics: VHyDRO achieves superior performance in Adjusted Rand Index (ARI), Change-point F1, and Mode F1 compared to HMM and ablation variants (no-support, no-mode). The "no-mode" baseline fails completely, confirming the necessity of discrete structure.

Experiment 3: Sparse Physical-Law Recovery

• Setup: Tested on four controlled hybrid systems with known sparse port-Hamiltonian equations.
• Findings: Given mode-pure segments, VHyDRO successfully recovers the active physical terms and coefficients.
• Metrics: VHyDRO achieves high Support F1 (0.978) and low relative coefficient error (0.006). In contrast, the "no-mode" ablation conflates regimes, leading to high errors, and the "no-sparsity" ablation fails to identify the active support despite reasonable vector-field prediction.

Significance and Claims

The paper claims that VHyDRO provides a necessary mechanism for stabilizing variational filtering in contact-rich environments where support mismatch is a primary failure mode.

• Mechanism Isolation: The authors emphasize that their results are "mechanism-isolation audits." They demonstrate that adding a defensive support carrier stabilizes filtering, which in turn enables coherent mode segmentation and sparse law recovery.
• Model-Relative Guarantees: The guarantees are explicitly model-relative. VHyDRO can only recover branches that are feasible according to the carrier model $P_t$; it cannot recover physical branches omitted by the model itself.
• Downstream Utility: The stabilized filter yields discrete states that are coherent enough to support downstream tasks like sparse physical law discovery, a capability that purely predictive baselines lack.
• Limitations: The authors modestly note that they do not claim closed-loop task-success dominance, real-robot transfer, or simulator contact-force ground truth recovery. The experiments focus on the intermediate mechanisms of filtering stability, segmentation coherence, and equation recovery. Theoretical guarantees for closed-loop control (Theorem 4) are presented as conditional downstream statements requiring a KL-calibrated rollout law, which is not numerically certified in the current experiments.

In summary, VHyDRO bridges the gap between scalable variational perception and interpretable hybrid dynamics by ensuring that the filtering process does not prematurely discard feasible contact branches, thereby enabling robust inference and physical law recovery in partially observed, contact-rich scenarios.