It uses an inpainting algorithm (adapted from image generation literature) to produce future actions that are consistent with the current trajectory. It's sort

It uses an inpainting algorithm (adapted from image generation literature) to produce future actions that are consistent with the current trajectory. It's sort of like warm-starting from a cached plan, although the plan isn't latent, it's directly in action space. Hopefully that answers your question -- there are many more details in the blog post and paper :)