Faculty Contact: Katie Byl
This project, which is an extension of the review conducted for Module 46, explores the use of robotics simulations for bipedal locomotion, specifically mujoco (http://www.mujoco.org/). We are interested in exploring movements that are "optimally" derived from deep reinforcement learning. For example, in the video (https://www.youtube.com/watch?v=hx_bgoTF7bs) at 1:28, the agent has learned to pump its arm in order to maintain balance. We would like to see if movements like this tell us anything about necessary mechanical movements that can be redone in a more intuitive way. This research combines a controls theory background with reinforcement learning, algorithm development, and multilayer networks.
- Summer 2018: Roman Aguilera