Researchers from UC Berkeley and Deepmind Propose SuccessVQA: A Reformulation of Success Detection that is Amenable to Pre-trained VLMs such as Flamingo