Data-Efficient Process Monitoring and Failure Detection for Robust Robotic Screwdriving

Screwdriving is one of the most prevalent assembly methods, yet its full automation is still challenging, especially for small screws. A critical reason is that existing techniques perform poorly in process monitoring and failure prediction. In addition, most solutions are essentially data-driven, thereby requiring lots of training data and laborious labeling. Moreover, they are not robust against varying environment conditions and suffer from generalization issues. To this end, we propose a stage and result prediction framework that combines knowledge-based process models with a hidden Markov model. The novelty of this work is the incorporation of operation-invariant characteristics such as screwdriving mechanics and stage transition graph, enabling our system to generalize across different experimental settings and largely reduce the required data and labeling. In our experiments, a system trained on M1.4x4 screws adapted with very little non-labeled data to three other screws (M1.2x3, M2.5x5, and M1.4x4) with widely varying tightening current, motor velocity, insertion force, and tightening force.

I. Introduction

Screwdriving is one of the most common assembly methods [1]. In the consumer electronics industry, hundreds of billions of tiny screws are assembled every year; however, fully automating this huge-volume assembly remains challenging, especially for smartphones [1], [2], [3]. Smaller screws require tighter tolerance and higher alignment accuracy [4]. Moreover, compared to the well-studied peg-in-hole problem (e.g., in [5], jamming diagrams for flexible dual peg-in-hole task were well-studied through large/small deformation stages, providing theoretical basis for control strategy design), screwdriving has more process stages and failure modes; most stages have complicated mechanics involving multiple contacts with highly variant discontinuous surfaces [6]. A system capable of online process monitoring, failure prediction and recovery is necessary for highly automated solutions [3]. However, existing work is still preliminary. Most of the previous work can only perform result classification given known failure modes, which alone cannot detect irreversible process failures [7] or unknown failures. Moreover, previous methods are not guaranteed to work when experimental conditions (e.g., screw sizes) change.