Previous work in RL for drone navigation has shown that using high-level representations of the environment, as opposed to lower-level representations such as RGB images, leads to more effective actions being taken by a RL model navigating an agent. We leverage Dynamic Scene Graphs (DSGs) and show that they are especially well suited as inputs into RL policies. Their hierarchical nature enables an agent to focus its search on a specific area, and storing explicit memory enables the agent to rule out previously visited areas. I work on creating a new interface for real-time DSGs to be input into RL models, and on training a new RL policy for these real-time scenes.

Participating in SuperUROP will help me extend knowledge I have gained in classes like 6.141 and my previous UROPs, and enable me to explore different fields in robotics. I hope to learn what goes into building scalable and reliable robotic systems, from a solid foundation in research, to testing and deploying products in the field. I will also be able to form meaningful connections with professionals and researchers in industry and academia.