GNNs for Knowledge Transfer in Robotic Assembly Sequence Planning

Abstract

Automated Assembly Sequence Planning (ASP) is a crucial task for robotic systems in manufacturing settings as it increases flexibility and efficiency. However, ASP is still largely performed manually, which can be time-consuming and prone to errors. The ASP process consists of two main steps: determining possible assembly sequences and confirming the feasibility of these sequences with the capabilities and restrictions of the target robot system. While several methods have been developed to automate ASP in recent years, they have limitations such as a lack of flexibility with regard to the properties of the assembly, a long training process, and a dependence on both positive and negative examples for feasibility detection. To address these issues, we propose a graph-based approach that divides the ASP problem into two tasks: Sequence Prediction and Feasibility Prediction. For the Sequence Prediction task, we model assemblies as graphs of part surfaces and apply a Graph Neural Network (GNN) to efficiently extract meaningful information from the input. An expert demonstrator guides us through the sequence prediction process, step-by-step, predicting which parts can be placed in their position at each state of the assembly. Our method is flexible and able to transfer knowledge between different assembly tasks. For Feasibility Prediction, we frame the task as an Anomaly Detection (AD) problem and use our GNN as a feature extractor to create latent representations for each assembly graph. We train a Normalizing Flows (NF) model to model the distribution of feasible assemblies and detect infeasible assemblies as Out-of-Distribution (OoD). Although our method performs better than a baseline one-class classifier, it still lags behind a binary classifier trained on both feasible and infeasible assemblies. However, it is not trivial to obtain a sufficient amount of infeasible assemblies for training. To better understand the limitations of our approach, we conduct ablation studies. Our approach requires only a few seconds to derive a feasible assembly sequence and could be integrated into future ASP frameworks to dramatically reduce planning time while using fewer computational resources compared to previous methods. To our knowledge, we are the first to use NF for graph-level Anomaly Detection. Overall, our method has the potential to significantly improve the efficiency and effectiveness of ASP in manufacturing settings.