Planning and learning goal-directed sequences of robot arm movements

摘要

Abstract: This paper describes two new types of neural networks. Firstly, a neural implementation of a resistive grid for path planning. Its advantages and limitations are discussed using the example of a 2-joint robot arm. Two of the limitations are the jerkiness of the movements and the inaccurate end-position. Secondly, as a solution to these two problems, a new type of sequence learning neural network is proposed. This 2-layer network can learn in one pass the sequence of movements defined by the resistive-grid. It uses RBF nodes with receptive fields centered on a sequence of starting positions in the configuration space of the arm, and with weights to the output layer being used to point to the next position in configuration space. The output layer uses a new type of node with output activity being equal to the activity-weighted average of the input weights. This network generates a smooth trajectory of the arm and an accurate end position. 1.