An error-learning neural network for the tuning of robot dynamic models

Abstract

The synthesis of accurate robot dynamic models is essential for the implementation of robust control strategies. However, robot dynamic models are coupled, highly nonlinear and of a continuously time-varying nature owing to external or internal disturbances. These disturbances can be caused, for example, by the different range of loads that must be handled by the end effector, physical wear and backlash of the motors, or simply by the adverse effects resulting from operation in a rapidly changing environment. Hence, a mathematically derived dynamic model might not turn out to be a good approximation of the actual system. This work presents an attempt adaptively to tune robot dynamic models in real-time situations. A neural network is used as a dynamic model compensator (DMC) to learn the error or the deviation between the model and the system. The neural network is implemented within a model-based framework to supervise and correct any discrepancy between the model and the system. The case is demonstrated by several examples.