Microsoft opens AirSim for AI training for UAV control

When moving, people and animals navigate relatively quickly, avoiding obstacles in an almost reflexive manner. In addition, if a person cannot immediately overcome another problem on his way - for example, open the door with an unusual handle, then in a few seconds or minutes of thinking the problem is solved and the door, as a rule, gives in. The next time this pen is no longer a problem. This, of course, is not only about doors and handles, but about solving such situations as a whole.

In addition, humans (as well as some animals) can predict which obstacle will appear in the next couple of seconds or even minutes. Seeing on his way a kiosk with newspapers, a person realizes that in 10-20 seconds he needs to go around. With robots (including unmanned vehicles and flying machines) all the more difficult. In order for them to be able to solve their problems on their own, they need to be trained. Microsoft, among other organizations, deals with this issue and is making some progress.

Now developers from Microsoft are creating a set of tools that will help third-party researchers and developers to train and test their own robots. The beta version of the software platform is available on the GitHub under an open-source license.
')
All of this is part of the Aerial Informatics and Robotics Platform research project (AirSim). As part of the project, researchers from Microsoft are developing software that serves as the basis for quickly creating third-party applications for managing drones and other gadgets. Also with the help of this system it is possible to train various devices that support self-learning.

Ashish Kapoor, the head of the project, hopes that the work of his team will be an incentive for progress in the field of “smart devices”, including ro-mobiles, product delivery systems and all other robotic devices. The goal of the project is to create systems that can work in the real world. This is the difference between the Microsoft project and the projects of other organizations, where the goal is to teach AI to control robots in an artificial environment with clearly defined rules. As an example, for example, board games.

“This is the next step in the development of AI, we are thinking about systems for the real world,” he says.

As an example, a situation where a drone needs to “explain” the difference between a shadow and a wall. It sounds simple, but in reality, if the copters begin to “understand” this difference, they will no longer be divided on obstacles. Or, in any case, the number of accidents will significantly decrease.

Interestingly, the corporation offers learning to lead in a virtual world that emulates the real world. Previously, this was not possible due to the limitations of hardware, and now, thanks to powerful graphics adapters and other components of computer systems, AI can be trained to drive the drone in a graphical environment that copies, for example, a forest or locality.

The positive point of this method of learning is the ability to change the conditions of the virtual world: time of day, weather conditions, relief. Another plus is the work of AI with virtual, rather than real, drones. And this is a cost savings, since drones in such tests are not needed, there are no accidents with the need to purchase new devices. Plus, there is no need to go to nature, change batteries, charge the drone. If the weather is bad - in the real world you cannot test the copter. And in the virtual world, the weather is always what the developer needs. Time losses are minimized.

Shital Shah, the main developer of the simulator, argues that AI training is possible only if high-quality graphics are created with precise environmental details.



In addition to the simulator, the Microsoft platform includes a software library that makes it easier for a developer to write his own code for the drones of two of the most well-known drones developers: DJI and MavLink. Usually, developers have to lose a lot of time analyzing the API when writing code for a specific hardware platform. Perhaps in the near future, Microsoft will add drones from other manufacturers.

According to representatives of Microsoft, AirSim allows you to train AI to predict the appearance of obstacles, as discussed above. This is necessary if a person wants to instruct AI to drive vehicles, copters, and other types of devices.

Microsoft employees have been working on creating their platform for less than a year. But before that, they took part in other projects. Some experts dealt with computer vision, others with AI or robotics. By combining all the experience of team members together, Microsoft received a new perspective development.

Now the team members are planning to create AI agents who can collaborate together and not compete. The same question, by the way, deals with the Google division, DeepMind. Previously, this unit conducted a detailed analysis of situations in which AI agents do not compete, but work together to achieve the same goal. Experts found out that the behavior of AI agents changes in accordance with the conditions and restrictions of their environment. If the rules assume a benefit in the case of aggressive behavior, the AI ​​will become more and more aggressive. Otherwise, the AI ​​will become increasingly more willing to cooperate with a partner or partners.

As for Microsoft, the company hopes that its new AI learning platform will allow the entire industry to grow at a faster pace. In addition, the company believes that it is necessary to begin to standardize the work of AI agents, introducing standards and rules describing the behavior of these agents in the real world.