I study CS at Cornell and do reinforcement learning research with Prof. Sarah Dean. I work on simulating and training robots.
Previously: software engineering at Bloomberg for two summers, and marine robotics research at the West Point Robotics Research Center.
A conditional diffusion model trained on ERA5 that generates realistic stratospheric wind for training balloon-navigation RL agents, extended with InfiniteDiffusion to effectively unbounded 4D wind fields. On a 12-metric realism benchmark it has 2.5× finer effective resolution than the Balloon Learning Environment's VAE.
An Isaac Lab locomotion environment for our six-legged robot. PPO from scratch learns to stand but never to walk; cloning a scripted tripod gait (BC + DAgger) and fine-tuning with PPO gives a policy that walks on command. Replacing 153 collision meshes with 19 primitives made simulation 2.4–4.5× faster.
A replication of Song et al. (IROS 2021) in the Flightmare simulator with PPO. The 8-gate circle track is solved two ways, with a hand-built curriculum and with an Optuna sweep. Along the way I found and fixed a collision check that treated the square gates as circles.
Search-and-rescue autonomy for unmanned surface vessels: a MOOS-IvP pipeline that runs YOLOv5 on the boat's live camera, detects people overboard and relays their GPS position. I demoed it live on the lake for NSA sponsors, trained the boats with DQN for Aquaticus (capture the flag on the water), and co-authored a SPIE 2023 paper.
My club's first runs on real hardware. Here we have a B601-RS arm picking up a roll of tape. Next up is running vision-language-action policies on it.
From Autoencoder to VAE trained on MNIST
From Monte Carlo estimates to REINFORCE.