Reinforcement Learning for Robotic Grasping in Construction

This document is a research article that presents the initial findings of a project focused on developing a Reinforcement Learning (RL) controller for robotic arms used in construction. The objective is to automate the stabilization of loads carried by cranes, thereby enhancing safety by reducing the need for human workers near oscillating objects. The paper discusses the challenges of integrating robotics into the construction industry, particularly due to the dynamic nature of construction sites. It outlines the project's first milestone, where an RL-controlled robot arm is tasked with grasping a cuboid at various positions in space. Initial experiments demonstrate the policy's robustness to different cuboid sizes and positions, although some performance issues in specific areas are noted. The methodology involves training the RL policy in simulation before potential deployment on real robots, with the ultimate goal of improving automation in construction processes and addressing safety concerns associated with manual stabilization of crane loads.