Announcement
Mirror project explores how AI agents can act on generated video plans
The post introducing Mirror says it collects video-to-action data by incrementally following a video planner.
TLDR
A post introducing Mirror calls it a “learning-from-experience” project that learns to follow video plans through trial and error. Its author claims Mirror significantly outperforms reinforcement learning in Minecraft without hand-crafted rewards or expert action annotations.
Combined views
9.6K
16 Sources, first seen ago
