Compositional learning provides a more structured, modular architecture to learn multiple tasks. My work is looking at how data provided to learn a specific set of tasks might be used to learn an entirely different task. This is an important area for continual learning for smart agents, since any real-life deployed agent would need to adapt and learn new tasks that wouldn’t be represented in the training dataset. I will be exploring how state representation algorithms can capture similar states between tasks and be used for leveraging data for separate tasks in the CompoSuite robotics benchmark.

I believe one of the most ambitious and world-changing visions is that adaptive smart agent operating in environments with humans. I am fascinated by this problem, and would like to understand what technical roadblocks need to be cleared and how I can contribute to this field. SuperUROP has provided me a way to pursue this research in a more substantial means, and I’m excited to be able to grow my awareness of academia’s approach to this goal.