Mirror's research is quite interesting! It uses a video model as a planner (fine-tuned with Cosmos3-Nano) to guide a video-to-action model, enabling the action model to perform V2A (video-to-action) self-play. Because the video model acts as the planner, the guidance process is re-planned based on the feedback state after the V2A execution. The V2A model essentially completes its self-play training without using reward signals or expert action annotations. This approach to V2A models is very similar to the ECHO paper from a few months ago (predicting the environment by training the model based on the real environment, where environmental feedback is the annotation for prediction).