I am a PhD student at the University of Washington, advised by Abhishek Gupta and Siddhartha Srinivasa. Previously, I received an M.S. in Robotics from Carnegie Mellon University, where I worked with Chris Atkeson and Katerina Fragkiadaki.
I am currently interning at NVIDIA. Previously, I interned with Amazon’s Frontier AI & Robotics team.
Automatically selects a minimal set of task-relevant keypoints for policy learning: the selection mask and a policy operating on the selected subset are jointly optimized end-to-end. The resulting keypoint-based policies remain robust to distractors, lighting changes and visual variation, and transfer across visually distinct scenarios.
Steers a pretrained diffusion policy to desired behaviors by running reinforcement learning over its latent-noise space, without touching the policy’s weights. Highly sample-efficient, DSRL enables real-world improvement of diffusion policies for robotic control and achieves state-of-the-art performance on simulated benchmarks.
Enhances the robustness of imitation learning by generating synthetic corrective labels. The trick is to leverage local continuity in the environment dynamics: a learned, locally smooth dynamics model synthesizes labels that guide the robot from out-of-distribution states back to the demonstrations — and for regions where the dynamics are discontinuous, it quantifies confidence and skips the unreliable ones.
Trains deep RL agents for dynamic fine manipulation without rigid surface support. Achieves nearly 100% success on the proxy task within 30 minutes of real-world interaction, and generalizes across disturbances, reset conditions, perception noise and object shapes.