Research Engineer · Astera Institute
My work spans learning in simulation, aligning perception and control interfaces, and deploying policies on physical robots. Previously, I received an M.S. in Intelligent Systems, Robotics and Control at UC San Diego, where I co-authored work on adapting offline world models in the real world.
I’m interested in how robots can learn goal-conditioned skills, reuse them across tasks, and adapt when the world changes.
Research interests: offline reinforcement learning, world models, goal-conditioned and reusable skills, and sim-to-real manipulation.
Pretrain a world model on offline robot data, then adapt it through online interaction. Test-time planning balances estimated return with model uncertainty to reduce extrapolation errors.
My contributions included simulation environments, simulation and task-transfer experiments, and corresponding real-robot experiments.
Value-Guided Action Transport studies residual action refinement toward higher estimated value, anchored to a behavior-cloned policy. I independently evaluated the method across twelve D4RL tasks spanning navigation, locomotion, and dexterous manipulation.
Trained camera-based insertion students with DAgger and deployed them on a high torque robotics arm. In simulation, I also trained specialist policies and a router for OpenArm stacking and insertion. Current research questions include local insertion search and transfer across starting conditions and part geometry.
Built a physics-filtered grasp pipeline to finetune a pretrained diffusion grasp generator, then integrated it with perception and a reaching controller co-developed with a teammate. The picking system uses a scripted skill graph to check outcomes and retry a failed grasp.
LOD / ACAIN 2024 · Springer LNCS 15510, 2025
Conference on Robot Learning · PMLR 229, 2023 · Oral presentation
NeurIPS 2021 Competitions and Demonstrations Track · PMLR 176, 2022
