Expert Demonstration Collection of Long-Horizon Construction Tasks in Virtual Reality

摘要

With the shortage of skilled labors in recent years, there is a pressing need for utilizing robots to perform repetitive and heavy construction tasks. Reinforcement learning (RL)-based robots become a promising solution because of their robustness and adaptability to unseen scenarios. However, long training time and complex reward design for these robots remain challenging. An effective solution is to collect expert demonstrations as inputs to better initialize policies of RL agents, or directly train inverse reinforcement learning (IRL) agents to recover reward functions. Therefore, this paper proposes a comprehensive virtual reality (VR)-based platform for expert demonstration collection. To show the effectiveness of our platform, a collaborative long-horizon construction task is implemented. We gathered 20 expert demonstrations as input to train a behavior cloning (BC) model. Results showed that the learned policy achieved reasonable success rates in completing the task, indicating the effectiveness of our demonstration collection platform.