research

A learned model of the physical world is only as trustworthy as the structure it respects. A network fit to trajectory data will often predict well over a short horizon while quietly violating the physics the real system obeys, and a controller that plans on that model inherits every violation, compounded over the rollout. My work builds the structure in rather than hoping it is learned: metriplectic flow matching that respects the thermodynamics of dissipative systems, fast and accurate system identification, and diagnostics for what neural PDE solvers actually learn. The same principle governs control — Lyapunov–barrier networks that enforce stability and safety jointly, model-predictive controllers that widen their safety margins when an ensemble of learned models starts to disagree, and imitation learning that reweights corrupted demonstrations so a robot does not inherit the mistakes in its data. New NSA-funded work carries the principle into perception, developing tools to attack and certify the symmetry-aware networks that encode 3D physical symmetries for LiDAR perception and robotic manipulation.

• ECCMetriplectic Conditional Flow Matching for Structure-Preserving Dynamics Learning2026
• NeurIPSMetriplectic Conditional Flow Matching for Dissipative Dynamics2025
• PLOS ONEForward Variable Selection Enables Fast and Accurate Dynamic System Identification with Karhunen-Loève Decomposed Gaussian Processes2024
• ICLRWhat Does a Neural PDE Solver Really Learn? A Residual-Spectrum Diagnostic2026
• RICODistributionally Robust Lyapunov-Barrier Networks for Safe and Stable Control Under Uncertainty2025
• OJ-CSYSHow Much Do Your Models Disagree? Adaptive MPC Safety from Ensemble Uncertainty2026
• ACCDensity-Ratio Weighted Behavioral Cloning: Learning Control Policies from Corrupted Datasets2026
• EAAIA Survey on Reinforcement Learning in Aviation Applications2024

Every method here eventually asks how far apart two probability distributions are. KL divergence is blind to the geometry of the underlying space, treating beliefs on adjacent states as no closer than beliefs at opposite ends of it — fatal for a safety argument, since close is the word every guarantee turns on. Optimal transport measures how much probability mass must move and how far, so error is expressed in the units of the state space: barycenters interpolate between disagreeing agents, trust regions are bounded in transport cost, flows contract at explicit rates. That last point is where the same geometry stops measuring and starts generating. In its dynamic formulation a Wasserstein distance is a velocity field carrying one distribution onto another — the object a flow matching model learns. Constraining that field to be metriplectic, to remain on a manifold, or to satisfy a temporal-logic specification is therefore not a second agenda but the same one: shaping transport paths so a learned simulator is stable in the very metric its guarantees are written in.

• L-CSSBlending Optimism and Pessimism via Wasserstein Barycenters for Continuous Control2026
• CDCWasserstein Stability of Contracting Flows: Effective Rates, Euler Self-Correction, and Noise Tightening2026
• L4DCWAVE: Wasserstein Adaptive Value Estimation for Actor-Critic Reinforcement Learning2025
• ICMLWasserstein-Barycenter Consensus for Cooperative Multi-Agent Reinforcement Learning2025
• NeuSLogic-Guided Vector Fields for Constrained Generative Modeling2026
• ICLRGeometry-Grounded Flow Matching on Compact Manifolds2026
• NeurIPSGeoReason: Step-Level Hallucination Detection via Hidden-State Trajectory Geometry2026
• ACCGeometry-Aware Decentralized Sinkhorn for Wasserstein Barycenters2026
• NeurIPSUnderstanding Reward Ambiguity Through Optimal Transport Theory in Inverse Reinforcement Learning2023

Most work on safe reinforcement learning assumes the constraint is handed to you in closed form. For a robot or an aircraft it rarely is: the specification is partial, expressed in temporal logic, implied by human corrections, or buried in a dataset of demonstrations of unknown quality. My work learns the policy and the constraint together — inferring unknown temporal specifications during training, refining behavior from human feedback, and correcting offline policies for constraints the data only implies. A recurring device is hierarchy: decomposing long-horizon tasks so that safety can be reasoned about at the level of abstraction where it is actually specified.

• OJ-CSYSConcurrent Learning of Control Policy and Unknown Safety Specifications in Reinforcement Learning2024
• ICAPSPolicy Refinement with Human Feedback for Safe Reinforcement Learning2023
• RSSImplicit Constraint-Aware Off-Policy Correction for Offline Reinforcement Learning2025
• MDPIMultilevel Constrained Bandits: A Hierarchical Upper Confidence Bound Approach with Safety Guarantees2025
• NeSyHierarchical Neuro-Symbolic Decision Transformer2025
• RLCFlow-Corrected Thompson Sampling for Non-Stationary Contextual Bandits2026

Trustworthiness requires more than average accuracy: a deployed system needs a defensible account of its own uncertainty, and a controller should become more conservative exactly when its models are least reliable. Conformal prediction offers distribution-free, finite-sample coverage guarantees, but standard formulations assume flat, exchangeable data — an assumption a physical system breaks routinely as it drifts into conditions its training data never covered. I extend it to settings with real structure — hierarchical data with group-level effects, and predictions living on manifolds — while preserving coverage, and study how conformal ensembles hold up when the distribution shifts online. Applications run from autonomous systems to clinical decision support.

• RAMConformal Prediction Across Scales: Finite-Sample Coverage with Hierarchical Efficiency2025
• Sci. Rep.A Hierarchical Conformal Framework for Uncertainty-Aware Length of Stay Prediction in Multi-Hospital Settings2026
• AAAIAdaptive Conformal Prediction via Bayesian Uncertainty Weighting for Hierarchical Healthcare Data2026
• ICMLGeometry-Aware Uncertainty Quantification via Conformal Prediction on Manifolds2026
• ForecastingHeterogeneous-Horizon Conformal Ensembles for Online Prediction Under Distribution Shift: An Empirical Comparison with Strongly Adaptive Methods2026