Both the abstract RNN and the MLP were only trained on data points whose sensory components were restricted to a two-dimensional manifold. The free parameters are the robot's distance to the center of the circle and the robot's orientation. Since it is not clear how a network reacts to points slightly outside its training domain, we have a look at the effect of a small change in the input of the forward model.
Let () be the transformation the network does on the sensory input. To each sensory input from the test set, in ten trials, a divergence was added. This divergence was distributed randomly and extended uniformly into the ten-dimensional sensory subspace. The magnitude of ranged between 0.0 and 1.0 pixels. The computation of () also requires a pair of velocities; in each trial, they were chosen randomly from the interval [-60; 60]. The results are in section 7.3.3.