Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos (Paper Explained)

#openai #vpt #minecraft 

Minecraft is one of the harder challenges any RL agent could face. Episodes are long, and the world is procedurally generated, complex, and huge. Further, the action space is a keyboard and a mouse, which has to be operated only given the game's video input. OpenAI tackles this challenge using Video PreTraining, leveraging a small set of contractor data in order to pseudo-label a giant corpus of scraped footage of gameplay. The pre-trained model is highly capable in basic game mechanics and can be fine-tuned much better than a blank slate model. This is the first Minecraft agent that achieves the elusive goal of crafting a diamond pickaxe all by itself.

OUTLINE:

0:00 - Intro

3:50 - How to spend money most effectively?

8:20 - Getting a large dataset with labels

14:40 - Model architecture

19:20 - Experimental results and fine-tuning

25:40 - Reinforcement Learning to the Diamond Pickaxe

30:00 - Final comments and hardware

Blog: https://openai.com/blog/vpt/

Paper: https://arxiv.org/abs/2206.11795

Code & Model weights: https://github.com/openai/Video-Pre-T...

Abstract:

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.

Authors: Bowen Baker, Ilge Akkaya, Peter Zhokhov, Joost Huizinga, Jie Tang, Adrien Ecoffet, Brandon Houghton, Raul Sampedro, Jeff Clune

Links:

Homepage: https://ykilcher.com

Merch: https://ykilcher.com/merch

YouTube: https://www.youtube.com/c/yannickilcher

Twitter: https://twitter.com/ykilcher

Discord: https://ykilcher.com/discord

LinkedIn: https://www.linkedin.com/in/ykilcher

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):

SubscribeStar: https://www.subscribestar.com/yannick...

Patreon: https://www.patreon.com/yannickilcher

Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq

Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2

Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m

Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n