Projects

TDV-LeWM planning results on OGBench-Cube

TDV-LeWM: Temporal-Difference Supervision for Latent World-Model Planning

Added a TDV-style motion encoder and temporal prediction loss to LeWM. Evaluated latent world-model planning on OGBench-Cube and improved cube success versus LeWM after accounting for no-op evaluation artifacts.

 Masked autoencoder reconstructions on STL-10

Masked Autoencoders on STL-10

Trained a ViT-S masked autoencoder from scratch in PyTorch on STL-10, then evaluated representation quality with linear probes, MLP probes, and last-block fine-tuning.

 Component labels on the vehicle

RoboSub Autonomous Underwater Vehicle Club

RoboSub focused on developing a fully autonomous underwater vehicle that could perceive its environment, estimate its state, and execute multi-stage tasks in a challenging 6-DOF setting. The system integrated computer vision, EKF-based estimation, PID control, Gazebo simulation, and ROS-based task logic into a complete perception-to-action pipeline.

 Component labels on the vehicle

CV Line Following and TDOA PCB Robot

Designed PCB that mounted microcontroller running FreeRTOS. Robot followed lines using CV, calculated TDOA direction, and was teleoperated via clap commands.