Answers are grounded in this paper and cite their sources.
Learning from human demonstrations has emerged as a promising paradigm in robotics. Especially in the context of disassembly tasks in circular factories, this would allow factory workers to teach robotics new skills without the need for robotics engineers or programmers, thus enabling flexible and scalable automation in modern manufacturing systems. However, collecting demonstrations only as a pre-deployment step is not flexible enough for rapidly changing production contexts. In circular manufacturing, end-of-life conditions are revealed during processing, and even in high-mix conventional manufacturing, variants and edge cases evolve faster than static offline datasets. We therefore argue for an integrated train-while-operate approach in which demonstrations are collected during productive operation, and policies are continuously updated. A key challenge of this approach is fluctuating demonstration quality caused by varying operator cognitive workload under real factory demands. This paper addresses this problem with a cognitive-workload-guided dataset-filtering framework for industrial learning from demonstration. We estimate demonstrator cognitive and physical workload during task execution and filter demonstrations collected under elevated workload. The framework is implemented and evaluated in a robot teleoperation scenario where cognitive workload is systematically varied to emulate realistic workload fluctuations and estimated using two sensing systems for comparison. Experimental results show that filtering the dataset with the proposed heuristics improves the performance of the learned motion models.
No takes yet. Share an insight, caveat, or question.
Wolf et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: