Openai video pre training (VPT): action learning based on watching unmarked online videos

title ：OpenAI|Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos（ Video pre training (VPT)： Action learning based on watching unmarked online videos ）

author ：Bowen Baker, Ilge Akkaya, Peter Zhokhov, Jeff Clune etc.

brief introduction ： This paper introduces an unsupervised sequential decision-making learning based on video data . Pre train noisy Internet scale data sets , It has been widely studied as a technology , It is widely used in training 、 Generic text 、 Model paradigms for images and other functions . However , For many areas of sequential decision making , Like robotics , Video games and computer use , Public data does not contain prior labels of behavior that need to be trained in the same way . The author uses the Internet scale to carry out pre training paradigm imitation learning on the sequential decision domain through semi supervision , Among them, agents learn actions by watching online unmarked videos . say concretely , The author shows how to use a small amount of labeled data , Train a sufficiently accurate inverse dynamics model , A large number of unmarked online resource data can be marked , Then train the general behavior a priori . Although using the local man-machine interface , The author proves that this behavior a priori has non trivial zero sample ability , And it can be fine tuned , And imitation learning 、 Reinforcement learning and exploration tasks cannot be learned from scratch through reinforcement learning . For many tasks , The author's model shows human level performance , The author is the first computer agent to report that can make diamond tools , Skilled human needs 20 More than minutes to complete the game .

Paper download ：https://arxiv.org/pdf/2206.11795v1.pdf