Natural creatures have evolved to optimize their body shape to be most suitable for survival. The innate body shape and the acquired control ability together determine their level of skills on different tasks. Modern reinforcement learning algorithms have proven to be successful in finding optimal control for robots on many challenging tasks. However, less attention is paid to optimizing the design of robots and developing a comprehensive benchmark for it. Previous work done in the Computational Design and Fabrication Group at MIT CSAIL has proposed a large-scale benchmark for co-evolving design and control of soft robots. The work spanned across multiple categories of tasks, including but not limited to locomotion, morphing, and manipulation on rigid, soft, multi-material, and varying terrains. This work is simulated in the Evolution Gym which is an environment created by the group to visualize co-optimization over concrete tasks, such as Walking’ or Climbing’ . In this SuperUROP project, we are going to expand upon existing work and extend the co-evolution to include multi-agent interaction.

I am participating in SuperUROP because I want to explore the intersections of my two interests, robotics and machine learning, through an advanced research project. I was initially exposed to these fields through ML and optimization coursework, as well as hands-on projects in the NEET Autonomous Machines thread. I am excited to be joining this group and look forward to being challenged and to furthering my knowledge in these areas.