The adoption of tele-operated and pre-programmed robots has already enhanced construction efficiency and safety, yet their application remains limited due to the need for expert remote control and challenges in adapting to dynamic environments.
To address these limitations, my study explores the potential of reinforcement learning (RL), particularly focusing on a novel approach that eliminates the need for manually designing reward functions. By integrating Generative Adversarial Imitation Learning (GAIL) with virtual reality (VR), my research enables robots to learn optimal actions from expert demonstrations and self-exploration. This innovative VR-GAIL model is applied to long-horizon collaborative construction tasks, involving a team of robots—an Unmanned Ground Vehicle (UGV) and two robotic arms—to perform complex operations such as transporting, picking, and installing window panels.
As shown in the figure, our approach mainly consists of three parts: (1) Designing two environments - a demo collection virtual construction environment and a Gym-like environment for RL training and evaluation. (2) collecting 20 expert-demonstrated trajectories for each sub-task, and (3) feeding the collected trajectories to GAIL while simultaneously training baseline PPO models.
The results are compelling: compared to the widely used Proximal Policy Optimization (PPO) model, our reward-free VR-GAIL approach achieved a 4.5% higher success rate across tasks and demonstrated superior adaptability as task complexity increased. These findings highlight the potential of VR-GAIL to improve RL performance while streamlining the development process by removing the need for complex reward function design.
This research represents a significant step forward in leveraging AI and robotics for smarter, more adaptive construction processes. I invite you to explore this exciting development in my full paper and video presentation. I look forward to contributing further to the transformation of the construction industry.
Feel free to reach out if you’d like to learn more or collaborate!