"Programming" by Teaching: Neural Network Control in the Manchester Mobile Robot

Paul Martin, Ulrich Nehmzow
1995 IFAC Proceedings Volumes  
The paper presents experiments conducted with the Manchester mobile robot FortyTwo on the application of reinforcement learning to robot control. An articial neural network, forming the core component of the controller presented here, associates incoming sensor signals with corresponding motor actions. Teaching the network is achieved by initially operating the robot under human control (i.e. the robot's behaviour is the result of training, not programming). The observed actions are used to
more » ... n the robot's associative memory, and after a very short training time FortyTwo becomes able to perform the required task autonomously. Experiments are presented in which data from sonar, infrared or vision sensors is associated with motor actions; and a suitable vision preprocessing system is discussed.
doi:10.1016/s1474-6670(17)46985-0 fatcat:frhww5zs7jatheljogt6tp724a