MIT's Mirror: video models as planners, agents learn from experience without rewards

MIT's Vince Sitzmann team released Mirror, a learning-from-experience project: an agent learns actions purely from the planning of video generation models. In Minecraft, with no rewards and no expert "video-action pair" labels, and paired with an oracle video planner, it crafts an iron pickaxe in about 30 minutes of game time. As video models improve, agent capabilities scale up alongside them.

Confirmed

• Mirror's core idea is that video models are getting better at "imagining" how to complete tasks, and the agent learns to execute those video plans; the authors explicitly state it is not RL but a form of behavior cloning that needs no "video-action pairs"
• Authors RyuHyunwoo and ericmchen1 argue embodied AI is overly obsessed with "actions": if video models scale like LLMs, they can generate "video plans" for any embodied task, and Mirror just needs to learn to execute them
• In an offline Minecraft environment, the video planner + IDM combination significantly outperforms the SOTA world-model agent Dreamer 4: roughly double the iron pickaxe success rate and about 8x on diamond acquisition
• Signs of emergent behavior: when dropped into a lava lake, the agent's escape rate over 30 attempts rose from 46% to 76%, learning in-context from its own history without any fine-tuning
• The authors predict that video models + LLMs open a new path for robotic learning-from-experience, possibly triggering an "arm-farm renaissance" that makes learning from robot experience viable

Why it matters

• Mirror offers a new paradigm beyond behavior cloning and RL: learning only to "execute video plans" sidesteps costly action annotation and reward design
• Tying agent capability to video model scale means progress in the underlying model translates directly into stronger embodied intelligence
• Emergent in-context learning behavior and clear advantages over the world-model route provide new empirical evidence in the embodied AI route debate

2026-10-20 ~ 2026-10-21 · 7 related posts

• [source] Mirror learns Minecraft skills from video plans with no rewards or labels in ~30 minutes — vincesitzmann · 2026-10-20
• Mirror author predicts an 'arm-farm' revival as video models make robot learning sample-efficient — vincesitzmann · 2026-10-20
• Mirror crafts an iron pickaxe in Minecraft with no rewards: video plans as robot policies — vincesitzmann · 2026-10-20
• Mirror video planner shows emergent in-context learning, boosting robot lava-lake escape rate from 46% to 76% — vincesitzmann · 2026-10-20
• [source] Video planners beat world models in Minecraft: 2x on iron pickaxe, 8x on diamonds — vincesitzmann · 2026-10-21
• [source] MIT's Mirror lets video models self-improve via environment interaction, first to craft an iron pickaxe in Minecraft — vincesitzmann · 2026-10-21

1 near-duplicate retellings: vincesitzmann