Intrinsic Robotic Introspection: Learning Internal States From Neuron Activations