research

Nearly every guarantee in the work that follows reduces to a question about probability distributions: how far apart are two agents' beliefs, how do you average them without destroying structure, how far may a policy move in a single update. Optimal transport answers these with a geometry rather than a heuristic. Treating policies and value estimates as distributions rather than points yields risk-sensitive value estimation, principled consensus among cooperating agents through Wasserstein barycenters, and trust-region updates that respect that geometry. A parallel line builds structure directly into generative dynamics models — metriplectic flow matching that respects dissipative physics, flow matching on manifolds, and temporal-logic guidance for constrained generation. The motivation is reliability rather than elegance: a learned dynamics model that violates conservation laws or drifts off the feasible set cannot support a safety argument downstream.

• 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
• ECCMetriplectic Conditional Flow Matching for Structure-Preserving Dynamics Learning2026
• NeuSLogic-Guided Vector Fields for Constrained Generative Modeling2026
• ICLRGeometry-Grounded Flow Matching on Compact Manifolds2026
• 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. In deployment 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 enforcing safety through learned Lyapunov–barrier certificates that come with stability guarantees. 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
• RICODistributionally Robust Lyapunov-Barrier Networks for Safe and Stable Control Under Uncertainty2025
• 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
• ACCDensity-Ratio Weighted Behavioral Cloning: Learning Control Policies from Corrupted Datasets2026
• ICAPSPolicy Refinement with Human Feedback for Safe Reinforcement Learning2023
• EAAIA Survey on Reinforcement Learning in Aviation Applications2024

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. I extend it to settings with real structure — hierarchical data with group-level effects, and predictions living on manifolds — while preserving coverage. A companion line turns disagreement among ensemble members into an online signal for adapting the safety margin of a model-predictive controller. Applications run from aviation autonomy to clinical decision support.

• OJ-CSYSHow Much Do Your Models Disagree? Adaptive MPC Safety from Ensemble Uncertainty2026
• RAMConformal Prediction Across Scales: Finite-Sample Coverage with Hierarchical Efficiency2025
• ICMLGeometry-Aware Uncertainty Quantification via Conformal Prediction on Manifolds2026
• AAAIAdaptive Conformal Prediction via Bayesian Uncertainty Weighting for Hierarchical Healthcare Data2026
• Sci. Rep.A Hierarchical Conformal Framework for Uncertainty-Aware Length of Stay Prediction in Multi-Hospital Settings2026

A learned controller can pass millions of random simulation trials and still fail catastrophically on the rare input that matters. Exhaustive verification is out of reach, and high-fidelity simulation is far too expensive to search with. I develop multi-fidelity falsification: methods that search broadly using cheap, approximate simulators and spend expensive high-fidelity evaluations only where a failure looks likely, with the fidelity level itself treated as a decision variable. Related work asks how much simulator fidelity is actually required to certify a given property, and develops stratified temporal logics for requirements that span scales. This thread underpins my NSF- and FAA-funded projects and was the subject of my AAAI New Faculty Highlights talk.

• IEEE AccessEfficient Counterexample Generation for Control Systems Using Multi-Fidelity Bayesian Optimization2026
• LogicsMulti-Fidelity Temporal Reasoning: A Stratified Logic for Cross-Scale System Specifications2025
• IEEE AccessBEACON: A Bayesian Evolutionary Approach for Counterexample Generation of Control Systems2024
• AI Mag.Exploring the Role of Simulator Fidelity in the Safety Validation of Learning-Enabled Autonomous Systems2023
• ECCFalsification of Learning-Based Controllers through Multi-Fidelity Bayesian Optimization2023
• AAAISafety Validation of Learning-Based Autonomous Systems: A Multi-Fidelity Approach2023
• arXivJoint Falsification and Fidelity Settings Optimization for Validation of Safety-Critical Systems: A Theoretical Analysis2023