This study presents a vision and proprioceptive data-driven robot control model for automating anchor-bolt insertion in construction. The model is designed to be robust under variable conditions such as lighting and hole surface states. It comprises a Spatial Attention Point Network (SAP) and a Deep Reinforcement Learning (DRL) policy, trained end-to-end to control the robot. The model demonstrates high effectiveness and efficiency, enabling task execution with higher success rates and shorter completion time. The approach can be easily applied to construction due to its high sample efficiency and short training time.
Publication date: 29 Dec 2023
Project Page: https://doi.org/10.1109/LRA.2023.3243526
Paper: https://arxiv.org/pdf/2312.16438
