Algorithmic Design of Autonomous Housekeeping Robots through Imitation Learning and Model Predictive Control

Abstract

Intelligent robots are more and more adopted into humans' regular life, from work to leisure. For instance, autonomous vehicles are running on public roads for testing, intelligent moving robots are deployed in hotel or museum lobbies to help customers. In this project, we will design algorithmic autonomous housekeeping robots to help people with housework. To enable intelligent and efficient motion planning that allows the robots to execute given tasks without colliding with humans or static obstacles (such as furniture at home), we use a combination of imitation learning and model predictive control (MPC). First, we will use MPC to generate and collect multiple optimal actions for randomly generated initial conditions of the robots, obstacles and target locations. Based on that, we use imitation learning to learn a policy network from the optimal policies generated by MPC. Moreover, we also adopt the concept of data aggregation (DAgger) to further improve the learning performance. The experimental results verify the effectiveness of our algorithms.