Modeling human intelligence and application to space object capturing

It is a great challenge for a robot In space to track and capture a free-flying object in the future space operations. In previous research, most of them employed model-based method which requires robot model in advance. However, it is difficult and time-consuming to obtain the precise mathematical model of robots. Moreover, the computer installed on the space robot is usually not so powerful due to the restriction of weight and volume, thus it is infeasible to computer dynamics parameters for real-time control. To facilitate the computation complexity, we present an approach based on human learning skill for tracking and catching objects. With human-teaching demonstration, the space robot is able to learn and abstract human tracking and capturing skill using an efficient neural-network learning architecture that combines flexible Cascade Neural Networks with Node-Decoupled Extended Kalman Filtering (CNN-NDEKF)[1]. The goal of skill learning is to obtain the most likely human performance from all the training examples and to transfer this skill to the space robot system by trained cascade neural network. We investigate the learning position trajectory in Cartesian space and position trajectory in Joint space respectively. Especially, learning position trajectory in Joint space is useful to avoid the complex inverse kinematics of space robot and to lower the computation cost. The simulation results attest that this approach is useful and feasible in tracking trajectory planning and capturing of space robot. The proposed approach provides a feasible way to plan the tracking trajectory of space robot by learning human demonstration skill. The learning is significant in eliminating sluggish motion planning and correcting a motion command that the operator may mistakenly generate. It will be found useful in various other applications, such as human action recognition in man-machine interfaces, real-time training, and agile manufacturing.