Artificial intelligence (AI) has come a long way in recent years. However, the technology is still far from perfect. While it excels at doing simple tasks and analyzing datasets, AI isn’t great at thinking flexibly like humans can. Researchers believe that this is the key to truly unlocking the technology’s potential.
The Allen Institute for AI believes that letting AI play hide-and-seek is the perfect way to teach it how to think, Digital Trends reports. This method helps virtual robots learn to carry out a wide variety of tasks.
Training AI bots in the real-world requires a lot of space. It’s also expensive since researchers need to build or purchase the robots. As such, the majority of AI training—especially in its early stages—is done in a virtual environment.
The Allen Institute has set up an extensive virtual playroom called AI2-THOR that consists of hundreds of apartments. Each is complete with kitchens, bedrooms, bathrooms, and living rooms and is accurate compared to real-world standards.
Inside the virtual rooms, AI bots take turns playing a game called Cache that’s very similar to hide-and-seek. One bot hides something within the environment. Then, the other is unleashed to try and find the hidden object. Over time, the bots get better at both tasks by learning which actions lead to a more favorable outcome.
Luca Weihs, a researcher from the Allen Institute, tells Digital Trends, “What makes this so difficult for Ais is that they don’t see the world the way we do… an AI starts from scratch and sees its world as a huge grid of numbers which it then must learn to decode into meaning.”
The training game helps teach the robots how to navigate and interact with their environment. It also reinforces skills like mapping, exploration, perspective, hiding, seeking, and object manipulation.
As noted, today’s AI isn’t great at performing tasks that aren’t well-defined. Likewise, an AI placed in a scenario that it wasn’t previously trained for is likely to struggle. That’s because we haven’t yet reached the point of artificial general intelligence (AGI). In other words, AI that can think, analyze, and adapt like a human.
It’s not hard to imagine why AGI technology would be useful. Think of something like a search-and-rescue robot. If guided by AGI, it could navigate any environment while searching for someone and adapting to obstacles along the way.
The Allen Institute’s work helps bring us closer to AGI, but the technology is still out of reach for now. To help AI take the next step, Weihs believes that letting it interact with the world—like it does when playing Cache—is crucial.
He tells Digital Trends, “I believe we will need to let agents learn the fundamental cognitive primitives we take for granted by letting them freely interact with their world. Our work shows that using gameplay to motivate AI agents to interact with and explore their world results in them beginning to learn these primitives—and thereby shows that gameplay is a promising direction away from manually abled datasets and towards experiential learning.”
This approach is worth monitoring as time goes on since it has the potential to dramatically reshape the way we think about and train AI systems.