I develop learning algorithms and control methods for sequential and real-time decision-making problems under uncertainty and constraints. My research draws on optimization theory to understand how policy structure and problem geometry influence what algorithms can learn and guarantee. I study reinforcement learning, policy optimization, and risk-tunable control, with an emphasis on performance, safety, stability, and convergence.

My work connects mathematical analysis to algorithm design and applications in autonomous systems and human–robot collaboration. Three questions guide this research: How can policies improve while respecting hard constraints? When does policy optimization admit global guarantees? How can real-time decisions account explicitly for uncertainty and application-specific risk requirements?

 Learning under hard constraints

Learning useful behavior requires both improving performance and respecting the constraints of the system. My research studies how to incorporate hard safety requirements into policy design and optimization, while establishing guarantees for the learning algorithm.

Safe policy optimization: I develop reinforcement learning methods that search directly over policies designed to satisfy specified safety constraints. We establish convergence and optimality results under the conditions of our analysis and demonstrate the approach in simulated quadrotor navigation. This work connects safe exploration with the optimization of the deployed policy. We expanded this work to design safe RL under unknown constraints and design learning-based safe control, which are currently under review.

Paper (AISTATS 2024), Repo

Safely learning infinite-horizon LQR: In a recent work, I study model-free policy optimization for linear quadratic regulation under coupled state–action constraints. This work examines how to design exploration over an undiscounted infinite horizon while respecting the specified constraints. We analyze the resulting optimization landscape and develop policy gradient methods with convergence guarantees under the assumptions of the framework. The aim is to connect constrained learning with the long-term performance and stability properties that motivate LQR.

Paper (NeurIPS 2026)

Multi-Agent Policy Learning for Optimal Transport (Ongoing)

I am developing learning methods for coordinating populations of agents through dynamic environments. This work investigates dynamically constrained optimal transport, where agents must move toward a target distribution while accounting for environmental flows.

My approach combines a learned shooting policy with Hamiltonian dynamics: the policy predicts initial control momenta from the collective agent configuration and flow descriptors, and the dynamical equations generate the trajectories. This structure connects policy learning to the underlying transport problem.

Optimization of structured policies

Policy structure can make learning more interpretable and connect the optimization problem to desired system behavior. I study how structured parameterizations shape the policy optimization landscape and enable algorithms with provable guarantees.

Policy gradient learning with PID policies: My collaborators and I formulate proportional–integral–derivative (PID) controller design as a policy optimization problem. We derive exact policy gradients and develop model-based and model-free learning algorithms. By establishing gradient-dominance properties, we obtain convergence and global optimality guarantees in the analyzed setting. We evaluate the algorithms on controlgym benchmarks, connecting optimization theory to a policy architecture widely used in practice. We have expanded this work to design learning algorithms over structure policies for unknown nonlinear dynamical systems while providing stability, optimality and convergence guarantees (under review). 

Paper (NeurIPS 2025), Repo

Real-time decision making under uncertainty

Real-time decisions must account for uncertainty in system dynamics, observations, and the surrounding environment. I develop optimization-based control methods that incorporate a designer-specified risk tolerance into safety-constrained decision making.

Risk-tunable safe control design: I use uncertain control barrier function (CBF) conditions to formulate chance constraints and develop sampling-based controller designs with probabilistic guarantees. The framework relates the chosen risk bound to the number of uncertainty samples required for a control design with desired safety satisfaction probability. The simulated quadcopter navigation illustrates how our safe control design adapts to the desired risk level.

Paper (CDC 2023)

Human–robot collaboration: We apply this approach to collaboration tasks with uncertain human position. We simulate and perform hardware experiments on a seven-degree-of-freedom Franka–Panda manipulator to characterize the relationship between the specified risk tolerance and task completion time. This work connects probabilistic safety requirements to measurable behavior on a physical robot.

Paper (TMECH/AIM 2025), Video

Safe Adaptive control for nonlinear dynamics: I also develop risk-tunable safe adaptive control methods for nonlinear systems. This work combines reference-model tracking with probabilistic CBF-based safety requirements, using a sampling-based design to tune controller parameters under dynamical uncertainty.

Paper (OJCSYS/CDC 2026)

Comparison of our Safe RL framework with with a safety-filter projection based safe RL