Researchers introduce Mirror, a self‑supervised video‑to‑action model

Experiments span Atari games, OGBench benchmarks, DeepMind Control suites, and the Minecraft BASALT challenge. Mirror outperforms leading learning‑from‑observation methods when paired with learned video plans and reaches near‑expert performance with oracle plans. In Minecraft survival, it is the first method to acquire an iron pickaxe solely through self‑improvement, and in the offline setting it sets a new state‑of‑the‑art result for diamond acquisition using expert labels only.

The results suggest that advances in video generative models can be leveraged to build increasingly capable embodied agents without relying on large expert action datasets or manually designed reward functions.