Learning to track human motion capture data has emerged as a promising method to control humanoid robots. However, simple motion tracking is not designed to handle real-world interactions and uncertainty, leading to brittle and/or dangerous behaviors. To address this, we aim to develop a whole-body motion tracking controller that responds to forceful contacts in a modifiable way (desired compliance level can be specified according to task). By training across a wide range of compliance levels, forceful interactions, and motion clips, we hypothesize the robot will learn appropriate responses for both expected and unexpected forces. As a result, we hope to demonstrate improved whole body manipulation task performance and safe responses to forceful disturbances.

I am participating in SuperUROP to gain hands-on experience in robotics and AI research. I am excited to learn more about reinforcement learning, modeling the physical world, and humanoid robots over the course of a longer-term project. I hope that this project will contribute towards safer, more functional robots.