Predictive modeling and planning of robot trajectories using the self-organizing map

Abstract

In this paper, we propose an unsupervised neural network for prediction and planning of complex robot trajectories. A general approach is developed which allows Kohonen's Self-Organizing Map (SOM) to approximate nonlinear input-output dynamical mappings for trajectory reproduction purposes. Tests are performed on a real PUMA 560 robot aiming to assess the computational characteristics of the method as well as its robustness to noise and parametric changes. The results show that the current approach outperforms previous attempts to predictive modeling of robot trajectories through unsupervised neural networks.