Robot Foundation Model Project

This project is part of the Cornell Human Robot Collaboration and Companionship (HRC^2) Research Lab.
I am currently working on improving, evaluating, and fine-tuning the pi-zero generalist AI model, a robot foundation model, for a variety of tasks such as sorting trash, folding laundry, etc. For those who don't know what a robot foundation model is, it is essentially a ChatGPT that can see and interact with its environment via cameras and robotic arms, allowing for generalized tasks to be done in the real world.

Some of my achievements:
Integrated π₀ generalist AI robot foundation model with Kinova Gen3 robotic arm using ROS 2. Improved success 50% → 90% (n=100) on a trash-sorting benchmark while cutting task completion time 2 min → 1 min via policy fine-tuning on custom high-quality rosbag2 datasets (dual RealSense cameras + Kinova Gen3 arm topics). Containerized the robotics stack (ROS 2, π₀ policy, drivers) for one-command, reproducible bring-up across machines. Designed and 3d-printed a 14g camera mount for the Realsense D435i camera.