DSPT: Disassembly Sequence Planning Transformer for Interaction Guidance in VR

The application of virtual reality technology in complex equipment disassembly training is widely used, and planning the disassembly sequence and interactively guiding the disassembly is an issue that requires in-depth research. Traditional methods based on physical collision detection are very accurate, but the computational efficiency is too low to meet the requirement of interactivity. In recent years, deep learning-based disassembly sequence prediction methods have emerged, which are fast in reasoning but suffer from inaccurate prediction of parts to be disassembled. In this paper, we propose a novel Transformer-based network, the Disassembly Sequence Planning Transformer (DSPT), to optimize the disassembly sequence for guiding users to disassemble objects in VR environments. First, we define Disassembly Sequence Features and Part History Features, along with their construction methods. Then, we introduce the parts-to-be-disassembled probability predictor based on a temporal-spatial score and propose a new loss function leveraging the temporal-spatial score to enhance the predictor’s performance. Experimental results show that our method achieves higher sequence accuracy and stepwise accuracy, both outperforming the state-of-the-art method. The results of the user study demonstrate that our method significantly reduces the disassembly task completion time and improves the usability compared to comparison methods.