The construction industry is facing an unprecedented shortage of skilled labor, creating an urgent need for innovative solutions to maintain efficiency and productivity. One promising avenue is the integration of robots to take on repetitive and physically demanding tasks. In particular, reinforcement learning (RL)-based robots stand out for their ability to adapt to dynamic and unpredictable environments. However, deploying these robots effectively presents challenges, such as the long training times and intricate reward design required to teach them how to perform tasks.
In my latest research, I explore how we can overcome these hurdles using expert demonstrations to enhance RL agents’ learning process. Specifically, my work focuses on a virtual reality (VR)-based platform designed to collect high-quality expert demonstrations. These demonstrations can then be used to initialize RL policies or directly train inverse reinforcement learning (IRL) agents to deduce reward functions.
To validate the platform, I implemented a collaborative, long-horizon construction task as shown in the figure below, and gathered 20 expert demonstrations. Using these inputs, I trained a behavior cloning (BC) model. The results were highly encouraging: the learned policy achieved a notable success rate in completing the construction task, demonstrating the potential of VR-based expert demonstration collection to streamline robot training processes.
This research paves the way for more efficient and adaptive robotics in construction, with the potential to address labor shortages and improve safety and productivity on job sites. Access the full paper by clicking here. Moreover, one demo trajectory collection video is also available here.
I’m excited about the opportunities this work presents for the future of construction technology and look forward to continuing to innovate in this field!