My research focuses on robot learning with generative models: learning multimodal distributions, remaining robust under distribution shifts at deployment time, and adapting behavior to test-time constraints. I combine imitation learning and reinforcement learning with experiments in simulation and on physical robots.
Autonomous surgical robotics is a central application of my PhD. Alongside developing learning algorithms, I build tools for collecting data, comparing policies, and deploying them on robots.
Generative imitation learning and trajectory planning
Understanding Multimodality in Generative Behavioral Cloning studies how latent information, deployment-time prior coverage, and the smoothness of action-space generative maps constrain multimodal behavior. The analysis is supported by synthetic navigation and physical-robot manipulation experiments, and examines the limited conditional multimodality of standard robotic simulation benchmarks. NeurIPS 2026, accepted.
Supervised Mixture-of-Experts for Surgical Grasping and Retraction develops a supervised Mixture-of-Experts policy for surgical grasping and retraction, with zero-shot transfer to ex vivo tissue and preliminary in vivo porcine evaluation. RSS 2026.
ANCHOR guides pretrained diffusion planners toward collision-free trajectories under test-time constraints, without retraining. Workshop paper accepted at STODY, NeurIPS 2026. I am a co-main author of the conference paper, which is under review.
Reinforcement learning and reward learning
I am interested in how policies improve through interaction and how human preferences can provide learning signals. My work explores learning from demonstrations and feedback, using simulation to develop policies and studying their transfer to physical systems.
Model compression and edge deployment
I investigate pruning and quantization algorithms to compress robot-learning models, reduce model size, and lower inference latency. My work explores deploying these models on resource-constrained edge devices and efficient AI hardware, while preserving policy performance.
Alongside these questions, I build reusable infrastructure for collecting data, comparing learning methods, and evaluating policies.