Data-centric and Goal-oriented AI for Robotic Repair Tasks

Robotic repair tasks in automotive manufacturing, such as spot repair, require heavy physical labor and knowledge of tools and products in order to make informed decisions about the repair strategy. Currently, they typically involve multiple manual steps, although automation pipelines exist that guide robotic repair using computer vision. Fully automating them requires sensible decision making based on past data and expected future consequences in terms of cost. We propose to apply machine learning techniques to this problem, show promising first results and discuss challenges for learning systems in this and related manufacturing processes. By using data-centric as opposed to model-centric AI techniques, we were able to improve the overall accuracy by about 6 %, and by using cost-sensitive learning, we were able to better guide model selection towards models that – at the same level in terms of accuracy and f1-score – shift unavoidable misclassifications to less severe outcomes (e.g. falsely treating repairable defects as irreparable). Therefore, in addition to the expected different costs of misclassification, we also consider the (non-monetary) decision preferences of endusers.