Deep Reinforcement Learning-based Task Assignment and Path Planning for Multi-agent Construction Robots

摘要

Recent developments in deep learning have enabled reinforcement learning (RL) methods to drive optimal policies for a sophisticated high-dimensional environment, which is suitable to overcome the challenges of implementing onsite construction robots, such as the dynamic nature of the construction environment and inherent complexity to solve the multiple decision-makers interacting simultaneously.In this research, we are trying to propose a systematic framework to adopt deep reinforcement learning (DRL) algorithms into onsite construction robotic applications (e.g., bricklaying platforms).This research has two main objectives: 1) Implement a multi-agent path-planning (MAPP) method for on-site robots that allow multiple mobile robots to navigate through the environment toward the assigned goal position and conduct the desired task logic while avoiding collisions, and 2) integrate the multi-agent task allocation (MATA) framework to solve complex tasks (e.g., laying bricks or delivering materials) through the cooperation of individual agents by assigning different tasks and roles to individual robots, which allows multiple robots to work simultaneously, just as how human workers act on a job site to make the best advantages of the productivity gains.