Goal-Conditioned Reinforcement Learning for Minecraft with Internet-Scale Video Pretraining

The article presents a method for training a goal-conditioned Minecraft model using internet-scale video data without explicit rewards, enabling it to achieve diverse and out-of-distribution goals. The model, Pan-4B, outperforms existing benchmarks like STEVE-1 and VLA in various tasks, demonstrating the effectiveness of pretraining on unlabeled video.

Comments

Read original article