Real world robotics data is currently tedious and time-consuming to collect. DART (Dexterous Augmented Reality Teleoperation) is an innovative approach that allows users to solve simulated tasks in augmented reality (AR) as if they were the robot, and to record their movements as training data for the robot to imitate. My project focuses on how to, given virtual demonstration data and a simulation environment, combine imitation and reinforcement learning to learn a successful control policy that is data-efficient and can best cross the sim-to-real gap. The implications of our research are a large scale robotics dataset and a novel, end-to-end training pipeline that can be used across the world by those without in depth robotics experience.

I am participating in SuperUROP to experience and contribute to a larger, more open-ended research project and to further my knowledge of deep learning (from 6.7960). I hope to explore modern approaches for robotic manipulation while making a meaningful impact on real-world robotics research through my project.