Adaptive Intelligence for Robot Navigation Efficiency with a Deep Reinforcement Learning-Based Cyber-Physical System

Abstract

The construction industry is exploring ways to enhance robotic efficiency in dynamic construction sites. Traditional robots rely on pre-set human programming, resulting in limited adaptability and efficiency. To realize the adaptive optimization of construction robots’ work efficiency, this study proposed and examined a Deep Reinforcement Learning-based Cyber-Physical System (DRL-CPS) paradigm. The physical components of the DRL-CPS contain an Unmanned Ground Vehicle (UGV) Robot which is mounted with a LiDAR sensor, preparing for distance capture in changing environments. The cyber components utilize Twin-Delayed Deep Deterministic (TD3) to train for Deep Reinforcement Learning (DRL) to achieve several task goals. The cyber training arena for TD3 is set by Robot Operation System (ROS) and the robot simulation environment. Then, the authors proposed a physical-layer experiment framework with the UGV robot and its accessory equipment in construction sites. After successful training of the TD3 model, the system was tested through two types of robotic navigation tasks: global and local navigation in different simulated construction environments. These tests, designed to assess the efficiency of the TD3 model against standard pathfinding algorithms and human operators, demonstrated the TD3 model’s superiority. The test results underscore the trained TD3 model’s advantages in reducing collision rates and minimizing time, achieving 40.6%, 31.0%, and 45.6% faster performance than the human tester in three unknown-map navigation tasks. The study demonstrates the potential of embedding adaptive intelligence in dynamic construction environments, paving the way for more efficient and intelligent robotic construction.