Learning a Motion Policy to Navigate Environments with Structured Uncertainty

Abstract

Navigating in uncertain environments is a fundamental ability that robots must have in many applications such as moving goods in a warehouse or transporting materials in a hospital. While much work has been done on navigation that reacts to unexpected obstacles, there is a lack of research in learning to predict where obstacles may appear based on historical data and utilizing those predictions to form better plans for navigation. This may increase the efficiency of a robot that has been working in the same environment for a long period of time.

This thesis first introduces the Learned Reactive Planning Problem (LRPP) that formalizes the above problem and then proposes a method to capture past obstacle information and their correlations. We introduce an algorithm that uses this information to make predictions about the environment and forms a plan for future navigation. The plan balances exploiting obstacle correlations (ie. observing obstacle A is present means obstacle B is present as well) and moving towards the goal. Our experiments in an idealized simulation show promising results of the robot outperforming a commonly used optimistic algorithm.

Second, we introduce the Learn a Motion Policy (LAMP) framework that can be added to navigation stacks on real robots. This framework aims to move the problem of predicting and navigating through uncertainties from idealized simulations to realistic settings. Our simulation results in Gazebo and experiments on a real robot show that the LAMP framework has potential to improve upon existing navigation stacks as it confirms the results from the idealized simulation, while also highlighting challenges that still need to be addressed.