Scalable and transferable learning of algorithms via graph embedding for multi-robot reward collection