Construction robots are challenging the traditional paradigm of labor intensive and repetitive construction tasks. Present concerns regarding construction robots are focused on their abilities in performing complex tasks consisting of several subtasks and their adaptability to work in unstructured and dynamic construction environments. Imitation learning (IL) has shown advantages in training a robot to imitate expert actions in complex tasks and the policy thereafter generated by reinforcement learning (RL) is more adaptive in comparison with pre-programmed robots. In this paper, we proposed a framework composed of two modules for imitation learning of construction robots. The first module provides an intuitive expert demonstration collection Virtual Reality (VR) platform where a robot will automatically follow the position, rotation, and actions of the expert's hand in real-time, instead of requiring an expert to control the robot via controllers. The second module provides a template for imitation learning using observations and actions recorded in the first module. In the second module, Behavior Cloning (BC) is utilized for pre-training, Generative Adversarial Imitation Learning (GAIL) and Proximal Policy Optimization (PPO) are combined to achieve a trade-off between the strength of imitation vs. exploration. Results show that imitation learning, especially when combined with PPO, could significantly accelerate training in limited training steps and improve policy performance.

翻译：建筑机器人正挑战着传统劳动密集型与重复性建筑任务的范式。当前对建筑机器人的关注集中于其执行包含多个子任务的复杂任务的能力，以及在非结构化动态建筑环境中的适应性。与预设程序机器人相比，模仿学习在训练机器人模仿专家执行复杂任务方面展现出优势，且后续通过强化学习生成的策略具有更强的适应性。本文提出一个由两个模块组成的建筑机器人模仿学习框架。第一个模块构建了直观的专家示教采集虚拟现实平台，机器人可实时自动跟随专家手部的位置、旋转与动作，无需专家通过控制器操控机器人。第二个模块利用首个模块记录的观测与动作，构建了模仿学习模板。该模块采用行为克隆进行预训练，并整合生成式对抗模仿学习与近端策略优化算法，在模仿强度与探索能力之间实现平衡。结果表明，模仿学习（特别是与PPO结合时）能在有限训练步数内显著加速训练进程并提升策略性能。