We talked about potentially doing something like this on a stream awhile ago using a VLM for visual feedback in runs not just rewards scores! https://www.krafton.ai/blog/posts/2026-04-03-prompt-to-policy/prompt-to-policy_en.html