research

I study safe and reliable autonomous systems at the intersection of control theory, robotics, and machine learning. My research develops data-driven, learning-based, and generative methods for safety verification, particularly for complex systems where exact computation is intractable and guarantees must be obtained from limited data drawn from complex distributions.

Using system symmetries, structural inductive biases, and compact latent representations, I aim to enable real-time, data-efficient safety verification that can be deployed in real-world robotic systems.

Research Overview

My research asks how robots can obtain meaningful safety guarantees with limited data and computation, across tasks where safety extends beyond collision avoidance to preventing object damage, spills, and hazardous interactions. I investigate diffusion models that learn compact latent dynamics and unsafe regions from labeled robot trajectories, together with safety certificates and filters for efficient verification in these learned spaces. A central goal is to understand the fundamental data requirements for safety and how physical priors, such as energy conservation, symmetry, and equivariance, can reduce them. By combining these inductive biases with task-dependent safety specifications, I aim to develop robot learning and control methods that are both data-efficient and verifiably safe.

Funding Projects

Amazon AI PhD Fellowship

This project aims to make embodied AI safer and more adaptable by developing safety certificates in the latent spaces of generative world models. It combines sampling-based reachability, and barrier functions with physics-informed and symplectic representations, action chunking and chaining, and adaptive exploration near safety boundaries to support robust generalization beyond imitation.

2026 JHU MINDS Fellowship

This project aims to develop a nonparametric framework and Monte Carlo methods for safety verification in high-dimensional dynamical systems using recurrence, recurrent sets, and control barrier functions. It investigates information-theoretic limits and adaptive sampling, while studying online certificate updates to reduce conservatism and support real-time safety verification as new observations become available.

Selected related work

Reachability and sample complexity

I study sampling-based reachability, including the roles of geometry, dynamics, and sample complexity. I also work on data-driven target reachability in Hamiltonian systems using symplectic inductive bias.

• On the Limits of Sampling-Based Reachability: Geometry, Dynamics, and Sample Complexity — Conference on Robot Learning, 2026.
• Symplectic Inductive Bias for Data-Driven Target Reachability in Hamiltonian Systems — IEEE Conference on Decision and Control, 2026.

Non-parametric methods for safety-critical control

My work explores safety-critical control and nonparametric safety verification.

• Safety-Critical Control via Recurrent Tracking Functions — American Control Conference, 2026.
• Recurrent Control Barrier Functions: A Path Towards Nonparametric Safety Verification — IEEE Conference on Decision and Control, 2025.

Learning, optimization, and energy systems

My earlier research includes smart predict-then-optimize methods with dependent data and shared storage for net zero energy buildings.

• Smart Predict-then-Optimize Method with Dependent Data: Risk Bounds and Calibration of Autoregression — arXiv preprint, 2024.
• Enabling Net Zero Energy Buildings With Shared Storage: A Cyber-Physical Perspective — IEEE Internet of Things Journal, 2023.

Full citations are available on the publications page.