End effector target position learning using feedforward with error back-propagation and recurrent neural networks

Concerns pattern-based system control by neural networks in dynamic situations. The neural network can be used to learn a sequence of mappings or space transformations. A 2-d.o.f. robot arm can be controlled to follow a sequence of steps from an initial position to a desired final position. The learning of a target position is understood as a dynamic control learning versus static learning. In this experiment the neural network is taught to guide robot arms from various initial points surrounding one arbitrarily chosen target point. The learning trajectories are straight lines on which the end effector moves. The closer the end effector is to the final point, the smaller angle increments are generated. The neural network is trained in such a way that these increments should approach the values close to zero in a finite number of steps. When this occurs, the implemented (computed) final position of a robot arm end effector should be as close as possible to the true final position. After training is finished, the feedback is added to the neural network and output signal increments are added to the input values. The neural network can generalize very well and lead the end effector to its final position from any initial position (different than those used for training) within the end effector's working envelope.<>