Microsoft says its high-fidelity simulated world lets developers generate enough data to train an AI system for autonomous flying.
Microsoft's latest artificial-intelligence project offers developers a set of tools to train robots, drones, and other gadgets to operate autonomously.
A key part of the project is AirSim, which according to Microsoft offers a very realistic open-source simulator for flying and crashing drones, to generate data for training autonomous robots and other vehicles. The Unreal Engine-based simulator is available on GitHub for anyone to use.
As with any AI project, one of the major challenges is accessing enough data to train a model for different tasks using a type of deep learning called reinforcement learning or trial and error. Google's DeepMind has had success using reinforcement learning to train its AI to master board games, like Go, as well as various 2D Atari games.
Microsoft says its high-fidelity simulated world will give developers a way to cheaply generate enough data to train an AI system for autonomous flying or driving in the real world. It would also allow developers to crash drones in a safe environment before unleashing one in the real world.
The simulator is part of a new Microsoft projected called Aerial Informatics and Robotics Platform. Microsoft's technical paper explains that the project aims to bridge the gap between simulation and the real world by using the latest graphics hardware and processors to mimic gravity, magnetism, and atmospheric conditions.
The simulator can also accurately render shadows, reflections, and weather-based effects, which can be helpful for any robot that relies on computer vision.
Currently the platform supports quadrotors, specifically DJI and MavLink-based drones, but Microsoft says modules for other vehicles can be added on.
According to Microsoft, the platform offers both cost and time savings.
"We can not only create various scenarios, but enact actions at a rapid rate, for example, hundreds of seconds of real world can be simulated in one second. Secondly, it enables carrying out and studying complex missions that might be time consuming and risky in real world. Finally, bugs and mistakes in simulation cost virtually nothing. We can crash a vehicle multiple times and thereby get a better understanding of implemented methods under various conditions."
Microsoft researcher Ashish Kapoor believes this project will offer AI developers a more practical way of training robots than using board games, which have well-defined rules.
"The aspirational goal is really to build systems that can operate in the real world. That's the next leap in AI, really thinking about real-world systems," Kapoor said.