Computer Science & Mathematics · UT Austin
I’ll be applying to PhD programs this fall, for Fall 2027 admission.
I'm an undergrad at UT Austin studying Computer Science and Mathematics, advised by Prof. Roberto Martín-Martín in the RobIn lab.
I work on the environments and data that robot policies learn from, simulation, synthetic demonstrations, and the reward and task design that shape what a policy ends up doing. I'm interested in how the world we build for a robot shapes what it learns.
Feel free to reach out! Always happy to talk about research.
TL;DR A simulator-agnostic plugin that tracks when a robot damages itself or surrounding objects, so safety can be measured. We use it to collect safer data, to train safer imitation learning and RL policies, and to evaluate VLAs.
TL;DR Mobile robots rarely stop at the exact spot where a demonstration was recorded, and visuomotor policies break when the camera moves. We rebuild the scene and the manipulated object with Gaussian splatting and re-render whole demonstrations from new base poses. Policies trained on this data still succeed more than 25 cm from the original pose, where every baseline fails.
TL;DR A benchmark that uses physics simulation to detect when a robot damages something during manipulation, giving one consistent way to score how safe its actions are. Policies trained with these damage metrics learn noticeably safer strategies.
TL;DR Before they can use a hammer, infants spend a lot of time banging objects on surfaces. We ran a study with ten infants to see whether the sound and feel of impact keep them doing it, then trained a simulated Franka to bang and reused that skill to learn targeted hammering. Reusing the banging skill as-is works sooner, while fine-tuning it ends up hitting the target more reliably.
TL;DR A person and a robot talk through a shared task, and either one can propose, accept, or turn down who does each step. In real-robot trials this raised task success and made the robot easier to work with.
TL;DR A robot learns a multi-step mobile manipulation task from one video of a person doing it. It splits the video into steps, maps each step onto its own body, and then practices on its own while avoiding unsafe actions.
Teaching assistant in the Robot Learning stream of UT’s Freshman Research Initiative. Mentored students on spatial transforms, kinematics, ROS, and PyTorch (FRI I), and supervised semester-long research projects on diffusion models, world models, and simulation benchmarks (FRI II).