MIT's Mirror lets video models self-improve via environment interaction, first to craft an iron pickaxe in Minecraft
vincesitzmann · x · 2026-10-21
Vincent Sitzmann and collaborators introduce Mirror, a self-supervised method in the paper 'Self-Improving Video-to-Action Models'. A video generative model acts as a visual planner while Mirror learns the video-to-action mapping purely from environment interactions—no human action data or hand-crafted rewards.
- The agent attempts to follow the visual plan, then trains on the resulting video-action pairs, incrementally improving.
- Outperforms leading learning-from-observation methods on Atari, OGBench, and DeepMind Control; near-expert with oracle plans.
- First method to acquire an iron pickaxe in Minecraft survival through self-improvement alone; sets a new SOTA on offline diamond acquisition.
- Early signs of emergent behavior: the video model remembers failed strategies and biases future sampling toward successful ones.
More from Embodied
- Stanford's FACT model uses human force signals to beat T-Rex by up to 60 points on dexterous tasks — stanfordnlp · 2026-10-21
- Stanford's T2-z lets robots probe hidden physics with a poke, lifting success from 34% to 73% — stanfordnlp · 2026-10-21
- Snap Hires Robert Downey Jr. as Chief Customer Experience Officer to Sell Specs AR Glasses Against Meta — adariostrange · 2026-10-21
- Upcoming paper claims big models can handle latency-critical real-time robotics — ManlingLi_ · 2026-10-21
- New project at the intersection of agents, embodied AI, and games gets a nod — ZeeshanZiaML · 2026-10-21
- NVIDIA open-sources Flora, a robotics data engine 125x cheaper than GPT-6 Astra — DrJimFan · 2026-10-21