Check out this cool new technique that’s making waves in the world of AI. So, you know how teaching AI agents can be a bit of a chore, especially with complex tasks like opening a kitchen cabinet? Well, researchers from MIT, Harvard, and the University of Washington have come up with a game-changer!

Usually, experts have to painstakingly design a reward system to motivate the AI to learn. It’s a bit of trial and error, and let’s face it, not the most efficient process. But these awesome folks have flipped the script. Instead of relying on expert-designed rewards, they’re tapping into the wisdom of the crowd.

Picture this: you want to teach a robot a task, and instead of experts, regular folks from around the world can chime in with their feedback. Yeah, it might get a bit messy with errors, but guess what? This new approach makes the AI learn faster than ever before!

Pulkit Agrawal, the brain behind this idea, says, “Our work proposes a way to scale robot learning by crowdsourcing the design of the reward function and by making it possible for nonexperts to provide useful feedback.” Fancy, right?

They call it HuGE (Human Guided Exploration), and it’s all about letting the nonexpert users guide the AI’s exploration. It’s like they’re leaving little breadcrumbs to help the AI find its way to the goal. Marcel Torne, the lead author, explains, “Even if the human supervision is somewhat inaccurate and noisy, the agent is still able to explore, which helps it learn much better.”

And get this – the AI doesn’t take the feedback too seriously. It uses it as a gentle nudge, exploring on its own and sending updates to the humans for a reality check. This method, tested on various tasks, proved to be faster and more effective than traditional approaches.

Imagine a future where your robot can learn tasks at home without you having to babysit it. With HuGE, the possibilities are endless!

The researchers are jazzed about refining HuGE to make the agent learn from other forms of communication, like language and physical interactions. They’re even thinking about applying it to teach multiple agents at once. Exciting times ahead! 🚀

Story Source:

Materials provided by Massachusetts Institute of Technology. Original story written by Adam Zewe and can be found here:  “New method uses crowdsourced feedback to help train robots.” ScienceDaily. ScienceDaily, 27 November 2023. www.sciencedaily.com/releases/2023/11/231127132237.htm