Learning to Ground Multi-Agent Communication with Autoencoders