Mirror assumes access to a pre-trained video model that, given the current observation, generates a video plan.
Self-improving agent learns video-to-action control from environment interaction alone
Mirror, requiring neither expert action labels nor rewards, outperforms prior learning-from-observation methods and is the first to reach an iron pickaxe in Minecraft via self-improvement.
Academic
Hyunwoo Ryu · Eric Ming Chen · Max Wilcoxson · Vincent Sitzmann
Massachusetts Institute of Technology
Research Digest··2 min read
The authors introduce Mirror, a self-supervised method that learns video-to-action mappings purely through environment interaction, without expert action labels or task rewards.
§