Enabling Grasp Synthesis Approaches to Task-Oriented Grasping Considering the <i>End-State Comfort</i> and <i>Confidence</i> Effects

摘要

Choosing a good grasp is fundamental for accomplishing robotic grasping and manipulation tasks. Typically, the grasp synthesis is addressed separately from the planning phase, which can lead to failures during the execution of the task. In addition, most of the current grasping approaches privilege stability metrics, providing unsuitable grasps for executing subsequent tasks. The proposed work presents a framework for high-level reasoning to select the best-suited grasp depending on the task. The best grasp is chosen among a set of grasp candidates by solving an optimization problem, considering the environmental constraints, and guaranteeing the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">end-state comfort</i> and the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">confidence</i> effects for the task, similar to human behavior. The framework leverages <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Generalized Bender Decomposition</i> to decouple the main non-linear optimization problem into sub-problems, thus presenting a modular structure. The method is validated with an experimental campaign using three different state-of-the-art grasping algorithms and three low-level motion planners in three different types of tasks: pick-and-place in a constrained environment, handover/tool-use, and object re-orientation. The experiments show that the proposed approach is able to find the best grasp, or at least one feasible, among the provided candidates for each task.