Mirror crafts an iron pickaxe in Minecraft with no rewards: video plans as robot policies
vincesitzmann · x · 2026-10-20
The Mirror team argues embodied AI has been over-obsessed with "action." If video models scale like LLMs, they can generate perfect video plans for any embodied task—Mirror simply learns to follow those plans through practice, with no rewards or action labels, crafting an iron pickaxe in Minecraft.
The companion thread predicts this "video model plans, behavior cloning executes" combo will revive arm-farms, as LLMs and video models make exploration sample-efficient enough.
More from Embodied
- Agility's Digit 5 drops running: leg redesign trades speed for safety and heavy lifting — chris_j_paxton · 2026-10-21
- Tesla Roadster's cold thruster system revealed, using the same actuators as Optimus legs — gregmushen · 2026-10-21
- Dev advocate builds a 6-wheel AI rover with Raspberry Pi 5, demoing at GTC Berlin — labeveryday · 2026-10-21
- Robotics evals show Gemini 4 far outperforms its general capability, says researcher — haider1 · 2026-10-21
- BrainLoRA adapts Qwen3-Omni to predict fMRI brain responses with 350x less compute — _akhaliq · 2026-10-21
- ePlane unveils full-scale 2.2-tonne eVTOL prototype e200X-PT01 built in Chennai — vaibhavbetter · 2026-10-21