You don't necessarily need neural networks for that, there are more specialized function approximations that learn faster and enable life-long learning for kinematics/dynamics and extrapolate better than NN which are more general purpose.