Intelligent trajectory planning and control for a robot manipulator using neural networks and evolutionary algorithms

Abstract

Recently neural networks, known as good universal approximators, have been widely used as powerful computational tool to effectively learn unknown nonlinear functions. It comes from an attractive idea that complex solutions can be obtained from learning with input-output data rather than explicit programming, which has made the neural networks emerge rapidly as a possible candidate to solve the complex problems. Due to such characteristics, interests in the neural-based applications to robotics have been much increased fro the last two decades. This paper deals with a trajectory planning for a robot manipulator. The planning is inherently a problem of multiobjective optimization. Especially, for given initial and final states, finding an optimal trajectory which satisfies a variety of objectives such as torque minimization, final state errors, obstacle avoidance, joint limitation, and so on, is very difficult. Moreover, since the planning is usually performed on the preestimated model dynamics, there exists a mismatch between the real optimal trajectory and the model-based trajectory. This paper systematically presents a trajectory planning method using learning capability of neural networks to overcome the mismatch. This paper is composed of 5 Chapters. After briefly addressing motivation and objective of the work and relationship between each Chapter, we explain several learning algorithms of the neural networks. Furthermore, to effectively deal with the underlying trajectory planning, multiobjective optimization using evolutionary algorithms(MOEA) and its usefulness are discussed. Based on Pareto optimality, several techniques to improve the performance of MOEA are proposed. And to guarantee good generalization of the neural networks, the network structure and learning conditions are empiricially studies. Based on the task specifications and necessary objectives, the optimal trajectory planning based on the model dynamics is performed with the neural networ...