A hybrid system for knowledge-based synthesis of robot grasps

Addresses the grasp synthesis problem arising in task planning for robotic dextrous hands. For this purpose, a hybrid architecture is proposed, which relies on symbolic and subsymbolic computations to exploit heterogeneous sources of knowledge, such as practical experience learned through experiments with a real device, heuristic rules gained from human observation, geometric reasoning and, when applicable, analytical results. After a preliminary discussion of representation levels and techniques, this paper describes the design of a tool for selecting the feasible grasps of a robotic hand under various situations and for ranking them according to task-oriented criteria. The interaction of a rule-based expert system with a neural network-based classifier provides support to both explicit reasoning and direct learning from experience. The features taken into account by the tool are object geometry, hand kinematic capabilities, task requirements, (in terms of both accessibility and robustness) and workspace constraints.