Hafner Dreamer agents reach Minecraft Diamond solve via offline world models before humanoid transfer

Hafner’s sequence of models—PlaNet, Dreamer 2, Dreamer 3, Dreamer 4—demonstrates progressive removal of real-world interaction. Dreamer 3 solved Minecraft Diamond from pixels. Dreamer 4 matched that result from recorded gameplay videos alone. The Technology Review profile confirms the same architecture now drives imported Chinese humanoids in unseen floor plans. DayDreamer already closed the loop on physical robots using the identical world-model loop.

Benchmarks show clear scaling: Dreamer 3 reached human-level Atari 2600 scores; Dreamer 4 removed all online samples. No comparable model-free method has matched the Minecraft result from offline data. The gap matters for robotics because sample collection on hardware remains orders of magnitude slower and costlier than simulation rollouts. Model-based planning therefore becomes the only feasible route to zero-shot deployment in homes or streets.

The profile understates the safety implication. Agents that plan inside an explicit world model expose their uncertainty estimates before action; model-free policies do not. This property directly addresses distribution shift in autonomous vehicles and manipulators. Hafner’s move to physical humanoids tests whether the same uncertainty-aware planning survives actuator noise and latency that video-game environments hide.

Next milestone is quantitative transfer: a single Dreamer-trained policy achieving 70 % success on novel real-world manipulation sequences within one hour of deployment, measured against a fixed task suite. Current evidence stops at qualitative video demonstrations.