Compositionality is one of the most important properties of natural language. Humans can
understand and parse sentences as well as reuse or compose new sentences in different scenarios. In this project, we will study how to encode compositionality into machine learning models for robot instruction-following by implementing and training different models to determine how well a robot is able to function given a sequence of commands. We will design a set of tasks and instructions suited for investigating the network’s ability to encode and decode compositional structures, both semantically and contextually. Our models will be tested on robots inside a drone simulation environment, from which we hope our work can be extended in future research to real world robot instruction following tasks.

“I want to participate in SuperUROP because it gives me the chance to focus on interesting technical research in a lab with incredible intellectual resources, while also developing my interest in machine learning. I have taken ML courses and done ML research which has inspired me to learn more and find practical applications of my knowledge. I am most excited to have my robot understand my ends, as well as improving my research skills.”