Control of a Multi-Joint Arm Using a Trial-and-Error Heuristic and a Neural Network

Abstract

We have developed a scheme that combines a gradient descent method, a trial-and-error heuristic and a neural network to control the reaching motion of a multi-joint arm. The arm can reach a target by avoiding obstacles in many cases. The basic control of the arm’s motion is made by a gradient descent on the potential field where a target and obstacles are represented as a low potential source and high potential sources, respectively. When the gradient descent gets stuck in a local minimum, the trial-and-error heuristic puts a high potential source, similar to the one for an obstacle, at the local minimum. Then the arm is reset to the initial- posture and begins the reaching motion again. The neural network learns good configurations of such high potential sources to fill in the local minima for various configurations of a target and obstacles from the past experiences of the trial-and-error. The neural network is shown to be helpful in reducing the number of trials not only in experienced cases but also in unknown cases, by guessing where to add the potential sources to prevent getting stuck.