Faster Confined Space Manufacturing Teleoperation Through Dynamic Autonomy With Task Dynamics Imitation Learning

摘要

Confined space manufacturing tasks, such as cleaning pilot holes prior to installing fasteners during aircraft wing assembly, currently require human experts to be inside ergonomically-challenging environments. Small rapidly deployable robots can substantially improve manufacturing safety and productivity. However, relatively rapid full automation remains elusive due to high-level of uncertainty in the environment, lack of cost-effective programming for low volume production, and difficulty of deploying adequate number of sensors in the confined space. Moreover, currently, teleoperation (remote human control of a robot via a force-reflection device) with typical levels of training and limited transparency of hardware is too slow for manufacturing applications, requiring experts to spend more time for each task to achieve the same cleaning quality. In this context, the main contribution of this article is to reduce cycle times for remote manufacturing by learning statistical dynamic autonomy from higher quality expert demonstrations in an ideal offline scenario. During the task, to keep cycle times low, the dynamic autonomy imitates the faster expert demonstrations when certain, and employs the slower human teleoperation when uncertain. A user study (n = 8) with an experimental robot platform shows that for the same cleaning quality, the dynamic autonomy reduces process completion time by 54.0% and human operator energy expenditure by 80.5% as compared with teleoperation without dynamic autonomy.