A Reinforcement Learning Based Approach for Conducting Multiple Tasks using Robots in Virtual Construction Environments

Abstract

—Construction robots are considered a promising solution for reducing onsite injuries and increasing productivity. One of the bottlenecks in deploying construction robots is solving the problem of robotic motion planning, considering the dynamic nature of construction sites. Specifically, current works in robotic motion planning for construction lack the generalization capacity for different tasks (i.e., a robot is generally optimized for a highly specialized task and fails to generalize when the task deviates slightly from its original form). In this paper, we proposed a reinforcement learning based approach for robotic motion planning using curriculum learning, which enables robots to conduct multiple construction tasks using a single trained agent. We tested our approach on three common construction tasks (ceiling installation, window installation, and flooring), resulting in an average success rate of around 80%.