To imitate or not to imitate: Boosting reinforcement learning-based construction robotic control for long-horizon tasks using virtual demonstrations: Fid-Bau Portal

To imitate or not to imitate: Boosting reinforcement learning-based construction robotic control for long-horizon tasks using virtual demonstrations

Huang, Lei / Zhu, Zihan / Zou, Zhengbo

Abstract

Abstract Construction robots controlled using reinforcement learning (RL) have recently emerged, showing higher adaptability and self-learning intelligence over pre-programmed and teleoperated robots. This work aims to train RL-based construction robots to learn long-horizon tasks with few exploration steps. We propose an approach that collects expert demonstrations in virtual reality, which are then added in the RL loop to assist with learning long-horizon construction tasks. Additionally, we utilize a hierarchical training strategy to generalize control policies to new policies that handle complex scenarios. For evaluation, we implement the approach for picking and installing window panels in simulation. In experiments, all 10 agents trained with virtual demonstrations delivered the task with success rates over 95%. Moreover, these 10 agents generalized their control policies and handled randomized window panels with success rates over 90%. The results confirm the effectiveness of our approach in boosting construction robots’ performance over long-horizon tasks.

Highlights • Built virtual construction site to collect expert demonstrations intuitively. • Developed an RL-based approach to learn long-horizon, sparse-reward tasks. • Proposed a hierarchical RL training strategy to generalize learned policies. • Empirically showed robot learns long-horizon tasks from demonstrations.