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-Training
内容中包含的图片若涉及版权问题，请及时与我们联系删除
评论
沙发等你来抢