A General Goal-Conditioned Minecraft Model
Summary
Pantograph reveals a family of goal-conditioned Minecraft models trained from internet-scale video, enabling zero-shot achievement of diverse, out-of-distribution goals. The Pan-4B model and friends are evaluated across 104 environments, showing strong performance in building, exploration, and mechanisms, while highlighting challenges like reward hacking and the benefits of scale. The work positions goal-conditioning as a scalable approach toward general robotics with future plans for broader data distributions and online RL.