Neural computing increases robot adaptivity

Abstract

The limited adaptivity of current robots is preventing their widespreadapplication. Since the biological world offers a full range of adaptive mechanisms working at different scales, researchers have turned to it for inspiration. Among the several disciplines trying to reproduce these mechanisms artificially, this paper concentrates on the field of Neural Networks and its contributions to attain sensorimotor adaptivity in robots. Essentially this type of adaptivity requires tuning nonlinear mappings on the basis of input-output information. After briefly reviewing the fundamentals of neural computing, the paper describes several experimental robotic systems relying on the following adaptive mappings: inverse kinematics, inverse dynamics, visuomotor and force-control mappings. Finally, the main trends in the evolution of neural computing are highlighted, followed by some remarks drawn from the surveyed robotic applications.