Research
I develop algorithms for safe autonomy — bridging control theory, machine learning, and robotics. My research unifies three tools: generative models for safe planning, data-driven control of nonlinear systems, and active inference under uncertainty.
Safe generative sampling
Diffusion models generate impressive plans and trajectories but offer no safety guarantees out of the box. I develop constricting control barrier functions that steer the reverse diffusion process to stay within safe regions — provably, without resampling or post-processing, and without modifying the underlying model.
- Prescribed-time safety guarantees via constricting tubes in the diffusion trajectory
- Applies to image generation, trajectory planning, and multi-robot swarm control
- No modification to the pretrained diffusion model required
Swarm control with dynamics-aware diffusion
Deployment on multiple Husky robots
Data-driven control via Koopman operators
Feedback linearization is a powerful control technique but requires an exact analytic model. I use the Koopman operator and its generator to learn the linearizing transformation directly from data — making interpretable, theoretically grounded controllers available for systems where no model exists.
- Data-driven feedback linearization with finite-sample guarantees
- Complete dictionary conditions for exact linearization from data
- Two IEEE Transactions on Automatic Control papers
Koopman-based nonlinear control demo
Active probing for intent inference
2024 · Honda Research Institute internship ·
Autonomous vehicles can't safely plan without knowing what other drivers intend to do. I developed an active probing framework that strategically selects actions to reduce uncertainty about other agents' behavior, while simultaneously planning safely under multimodal predictions.
- Closed-form Wasserstein risk metric for multimodal Gaussian mixture predictions
- 44% smoother trajectories vs non-probing baseline in MetaDrive simulation
- Risk-aware thresholding prevents dangerous probing actions
Active probing in highway merging scenarios
Coverage control of ground robots
Distributed algorithms for optimal area coverage using teams of ground robots with OptiTrack mocap. Consensus protocol for oscillator-based controller.
Behavior-based attack detection in dynamical systems
Direct vs indirect methods for detecting attacks using behavioral analysis. Established a tradeoff between system identification and direct detection.
Heterogeneous online learning
Fusing multiple algorithms for online learning with heterogeneous sensor arrays arriving at different rates.