An artificial intelligence can make its own internal map of its environment to navigate as animals and humans do, even though it can鈥檛 see. The map can be transferred between different AIs and could be useful for improving real-world navigation of robots or improving their skills at manipulating objects.
When we encounter a new environment, our brains produce an internal map that is unique to us, unlike a map you might see on Google, and we use it for finding the same things again or taking shortcuts.
Now, at the Georgia Institute of Technology in Atlanta and his colleagues have trained an AI agent with no visual sensors 鈥� it is effectively blind 鈥� to navigate in an environment and find a destination. They have shown that it appears to generate an internal 2D map of its environment.
Advertisement
These blind agents existed in a virtual world called Habitat and only had one type of sensing, known as egomotion, which told them how far they had moved and in which direction. The agents were then told to find their way from one point to another, through an obstacle-strewn room, with their only guide being how close they were to their destination.
The agents could find a route to the destination that was 1.59 times longer than the shortest path. Similar agents that can see normally take a route that is 1.19 times longer than the shortest path.
When the memory of blind agents was transplanted into those that could see, they could find an even shorter path. When a blind agent鈥檚 memory was removed, it lost the ability to navigate effectively.
Some previous AI models have navigated environments without being given any external maps, but because the workings of AIs are opaque and they figure out how to do things themselves, it was unclear whether they were using self-made internal maps like humans and other animals do or instead using visual cues like landmarks.
Wijmans and his team now think the agents are building and relying on internal maps, because their blind agents can鈥檛 have been using visual cues.
The emergence of internal maps could be useful, says 鈥� at the Allen Institute for Artificial Intelligence in Seattle, Washington. 鈥淚f you can teach [an AI] to navigate and that results in the emergence of a mapping capability, then you could probably use this model, or fine-tune this model, for other tasks which also require mapping.鈥�
This could also extend to non-navigation tasks, says Kembhavi, such as generating 3D schematics of an object after picking it up and using these for further precise object manipulation.
Reference:
