Programming robots and designing optimal robot controllers are complex problems. Models of robots and their environments become very complex for real-world applications. Often, the optimal controller is nonlinear. Synthesising a non-linear optimal controller off-line is often infeasible. Several neural learning approaches attempt to solve these problems. This article presents several examples and experiments in robotics, illustrating advantages of neural control. In particular, two assembly experiments are described in detail. In the first experiment, a robot must insert a peg into a hole. In the second experiment, a robot must insert the base part of an electric switch into a fixture. The presented experiments use various neural net architectures: the backpropagation net, the radial basis function net and the cascade correlation net.