Hardware Conditioned Policies for Multi-Robot Transfer Learning