In this paper, we propose a new method to learn a multi-fingered hand grasping posture with little knowledge about the task and few sensing capabilities. The developed model is composed of two stages. The first is dedicated to the finger inverse kinematics learning in order to provide the fingertip-desired position. This function is fulfilled by modular neural network architecture. Following the concept of reinforcement learning, a second neural model dealing with noisy sensing information is used to search the space of hand configuration. Simulation results show a good learning of grasping postures with five fingers and different noise levels.
Rapid explorative direct inverse kinematics learning of relevant locations for active vision
