On how robotics can benefit from dense visual representations, neural radiance fields, and large language models