Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos
This work addresses the challenge of learning from unlabeled data in sequential decision-making for domains like gaming and robotics, offering a novel method to leverage abundant online videos, though it is incremental in applying pretraining paradigms to new modalities.
The authors tackled the problem of training behavioral priors for sequential decision domains like robotics and video games by extending internet-scale pretraining to unlabeled online videos, using a semi-supervised imitation learning approach with a small labeled dataset to label large-scale Minecraft gameplay videos, resulting in agents that achieve human-level performance on tasks such as crafting diamond tools, which take humans about 20 minutes (24,000 actions).
Pretraining on noisy, internet-scale datasets has been heavily studied as a technique for training models with broad, general capabilities for text, images, and other modalities. However, for many sequential decision domains such as robotics, video games, and computer use, publicly available data does not contain the labels required to train behavioral priors in the same way. We extend the internet-scale pretraining paradigm to sequential decision domains through semi-supervised imitation learning wherein agents learn to act by watching online unlabeled videos. Specifically, we show that with a small amount of labeled data we can train an inverse dynamics model accurate enough to label a huge unlabeled source of online data -- here, online videos of people playing Minecraft -- from which we can then train a general behavioral prior. Despite using the native human interface (mouse and keyboard at 20Hz), we show that this behavioral prior has nontrivial zero-shot capabilities and that it can be fine-tuned, with both imitation learning and reinforcement learning, to hard-exploration tasks that are impossible to learn from scratch via reinforcement learning. For many tasks our models exhibit human-level performance, and we are the first to report computer agents that can craft diamond tools, which can take proficient humans upwards of 20 minutes (24,000 environment actions) of gameplay to accomplish.