Learning-based multi-path reconstruction for 3D object-centric robot motion planning

Autonomous path planning for interaction with complex 3D shapes is a critical task in a variety of industrial processes such as spray painting, polishing, and welding. In such scenarios, a robot is typically required to generate multiple end-effector pose paths matching the surface geometry of the input objects. Recent work demonstrates the feasibility of generating path segments for objects with complex shapes via 3D deep learning techniques leveraging expert demonstrations. In this context, a relevant open problem is the combination of such segments into a viable set of long-horizon paths, as generated segments are neither assigned to different paths, nor ordered within each path. This thesis targets this problem by applying and evaluating recent learning-based approaches. In particular, it investigates: (1) The performance of DBSCAN, a density-based clustering algorithm used for path enumeration that achieves satisfactory results on objects with simple shapes, yet fails on objects with more complex surfaces; (2) The limitations of a heuristic algorithm for the open Travelling Salesman Problem (TSP), which is employed for sorting segments within sequences. This method is highly sensitive to the initial segment, which is critical for good sequences. To that end, two ad-hoc heuristics for selecting the starting segment are introduced; (3) The predictive performance of a segment sorting method based on Graph Neural Networks (GNNs). This method formulates concatenation as an edge prediction problem in an oriented graph, where each node represents a segment. It is demonstrated to be an efficient method for reconstructing the ordering and is robust to noise; (4) The feasibility of a transformer-based method for single-step multipath prediction. This method aims to address clustering and concatenation problems jointly, in contrast to previously mentioned approaches. Preliminary results on a single object category suggest the applicability of this method. Extensive experiments on a real-world dataset collected in an industrial scenario for a spray painting task on multiple object categories suggest how learning-based approaches can be effectively used to reconstruct long-horizon paths which preserve task-specific patterns given expert demonstrations. Furthermore, learning-based approaches improve robustness to noise when compared to traditional heuristics for TSP problems.