Learning a Motion Policy to Navigate Environments with Structured Uncertainty

This project focuses on reducing the distance a robot has to travel when repeatedly moving between destinations in an uncertain environment.

The problem is motivated by robots being deployed in offices, warehouses, or hospitals to repeat mundane tasks such as moving materials. However, there may be unexpected changes such as doors being closed or forklifts blocking a route. The goal is to improve their efficiency by learning to predict changes in the environment based on current observations. We have developed novel algorithms to reduce the average distance traveled as more tasks are executed, and a framework to implement these algorithms on real robots.