The researchers claim their method could be used to train artificial intelligence to perform other tasks. First, it could be used for robots that use a keyboard and mouse to browse websites, book flights, or buy groceries online.But in theory it could be used to train robots to perform physical, real-world tasks by replicating First-person videos of people doing these things“It makes sense,” Stone said.
Matthew Gudzial of the University of Alberta in Canada uses video to teach AI the rules of games like Super Mario Bros., but he doesn’t think that’s going to happen anytime soon. Actions in games like Minecraft and Super Mario Bros. are performed by pressing buttons. Actions in the physical world are much more complex for machines and harder to learn. “It unlocks a whole bunch of new research questions,” Gudzial said.
“This work again demonstrates the ability to scale up models and train on massive datasets to achieve good performance,” said Natasha Jaques, who works on multi-agent reinforcement learning at Google and UC Berkeley.
Large data sets the size of the Internet will certainly unlock new capabilities for AI, Jaques said. “We’ve seen this over and over again, and it’s a good approach.” But OpenAI is a big believer in the power of big data sets, she says: “Personally, I’m more skeptical that data can solve anything. “
Still, Baker and his colleagues think collecting more than a million hours of Minecraft video will make their AI better. It’s probably the best bot to play Minecraft so far, Baker said: “But with more data and a bigger model, I want it to feel like you’re watching a human being play the game, rather than a bot trying to mimic a human being.” Baby AI.”