Can Artificial Intelligence Understand Maps

Maps are constructed by human beings. Through the medium of map, people can understand space through the expression of space. Symbols and descriptive languages in maps are universal, so people can use maps to navigate in familiar or unfamiliar places. This is an important skill, and the military is still using compasses and maps to train navigational skills.

Can Artificial Intelligence Understand Maps

Imagine if AI could read and track maps in the same way.

Cross-Perspective Policy Learning of Street Navigation

DeepMind, a British-based AI company that shares Alphabet with Google, published a report on June 13 entitled Cross-View Policy Learning for Street Navigation. The report is an evaluation of an experiment designed to test whether AI can train navigation by using maps instead of drawing a lot of routes for it like autonomous navigation tools.

The report's authors point out that "goal-driven Street navigation agents have so far been unable to move to invisible areas without extensive retraining, and relying on simulation is not an extensible solution. Because aerial images are easily accessible globally, we recommend that a multi-modal approach be trained for ground and air views, and then benefit from it." Use aerial view observation to transfer the ground view method to the invisible part of the city.

The difference between this method and the automatic driving vehicle navigation algorithm training method is that it increases the map from the top to the down as a reference point, and uses a completely new information in the navigation model. Traditional autonomous navigation algorithms run on images already drawn, which are scanned data stored in the robot brain to teach it how to drive on familiar roads.

The researchers wrote: "The discovery that humans can quickly learn about a new city by reading maps inspires us to try to integrate similar top-down visual information into the training process of navigation agents to help them spread to streets we have never seen before. Instead of using maps drawn by humans, we chose aerial imagery, which is easily accessible around the world. In addition, once people are familiar with an environment, they can not use maps. This diversity of human beings has inspired us to train flexible RL agents, which can be executed using either first-person views or top-down views."

Testing process

To test this new method, AI simulates operations in an unfamiliar place, choosing to go forward, turn left or right, turn left or right to a greater extent. AI uses panoramic street view images to navigate the surrounding environment, sometimes using map views to run. In experiments, AI using map views is easier to navigate successfully to where it needs to go than AI without map views.

For unfamiliar areas where accurate data can be used, it is useful to use such an algorithm to guide military vehicles. Although some high-quality satellite imagery can be obtained almost everywhere on Earth, street-level imagery, especially street-level imagery rendered by lidar, and image iteration by driving algorithm, is much less. Compared with the path simulated accurately in advance, the task of teaching robots to walk in space based on their own sensors and maps is much less, and more matches the current capabilities of most remote control and autonomous systems. Essentially, it only realizes the function of human choosing path points on the GPS drop-down menu. The only advantage is that AI can navigate by itself and is independent of GPS design. DARPA is looking for artificial intelligence that can provide military applications based on precise data, and DeepMind's map learning and navigation tools may meet similar needs.

DeepMind's researchers told an interesting story at the beginning and end of the paper: how Ernest Shackleton's team first used a map and then, when it failed, used the memory of the map to map a path out of the ruins.

For Shackleton, map memory is a tool to escape the unknown. For the Pentagon, AI, which can read maps, may be a way to challenge the unknown.