Learning-based robot assembly method for peg insertion tasks on inclined hole using time-series force information

This paper presents a novel method for learning force-aware robot assembly skills, specifically targeting the peg insertion task on inclined hole. For the peg insertion task involving inclined holes, we employ one-dimensional convolutional networks (1DCNN) and gated recurrent units (GRU) to extract features from the time-series force information during the assembly process, thereby identifying different contact states between the peg and the hole. Subsequent to the identification of contact states, corresponding pose adjustments are executed, and overall smooth interaction is ensured through admittance control. The assembly process is dynamically adjusted using a state machine to fine-tune admittance control parameters and seamlessly switch the assembly state. Through the utilization of dual-arm clamping, we conduct key unlocking experiments on bases inclined at varying degrees. Our results demonstrate that the proposed method significantly improves the accuracy and success rate of state recognition compared to previous methods.