Virtual Reality-Based Expert Demonstrations for Training Construction Robots via Imitation Learning: Fid-Bau Portal

Virtual Reality-Based Expert Demonstrations for Training Construction Robots via Imitation Learning

Huang, Lei / Cai, Weijia / Zou, Zhengbo

The construction industry faces challenges of skilled labor shortage, low productivity, continuous cost and schedule overruns, and unsafe working conditions for workers. The application of construction robots is a potential solution to alleviate these problems by providing higher productivity, quality, and a safer working environment. Existing construction robotic solutions, such as bricklaying robots and grid drawing robots, are typically programmed to follow specific sequences of instructions designed beforehand. Recent advancements in robot control algorithms like Reinforcement Learning (RL) have enabled robots to adapt to unseen scenarios by learning sequences of optimal actions from videos of expert demonstrations. However, these demonstrations need to be collected on real construction sites, which can be costly and potentially dangerous to experts, especially when multiple demonstrations are required. In this study, we propose a novel approach that leverages Virtual Reality (VR) to collect expert demonstrations, allowing the demonstrator to illustrate the procedure of a construction task using handheld VR controllers. During demonstrations, direct parameters including the robotic arm’s joint states, positions, and orientations of the object to be manipulated are extracted from the virtual environment to generate a control policy that imitates the behavior of the expert. We implemented the proposed approach for the task of window installation as validation. The control policy was learned and later applied to a robot arm in a virtual environment. Results show that for all 10 testing cases, the control policy could successfully generate actions given the observed states and lead the robotic arm to first pick up the window and then install it at the target location. These results confirm the effectiveness of the proposed approach in providing virtual demonstrations for the robot to learn a control policy, which eliminates the need for on-site demonstrations from experts, avoiding potentially unsafe scenarios.