This project will focus on finding novel state representations for deformable objects and using these representations to improve downstream manipulation performance. The initial focus will be on a selection of fruit and vegetable items that must be prepared for cooking. Such state representations should have desirable properties, including being entirely obtained through vision, being task-agnostic, and being fully learned through self-supervision. These representations should also be efficient with respect to the information required for achieving manipulation tasks; preparing a salad does not require detailed information about the state of each of its leaves. Can we obtain both efficient and task-agnostic representations?

Through this SuperUROP, I want to gain research experience to balance the industrial experience I acquired from my past internships. I spent nine months working on robotic manipulation at Ocado Technology and am excited to help advance the field through research. I hope to expand my robotic manipulation knowledge, gain research experience for my future PhD, and publish a paper on my research.