The technology might be young, but we've already seen autonomous drones do some truly amazing things. Behind those impressive tasks lies a lot of work – not only on the part of the drone builders, but from the drones themselves as they seek to improve through machine learning. Microsoft has just launched a simulator platform where robot-makers can train their machines and gather useful data, all without fear of destroying expensive hardware.
The system is called the Aerial Informatics and Robotics platform and the idea is that it can help minimize the trial-and-error costs involved in developing useful robots. This means providing developers with a way of debugging their systems during the training phase, without needing to unleash a physical, unpredictable robot.
As for what it looks like, the graphics are certainly impressive and realistic enough to give a visual sense of real-world conditions. Microsoft has begun by implementing a quadcopter on the platform, and in the video below you can see it take flight on an urban street and navigate power lines, poles and trees. In addition to first-person view from the drone, real-time object-depth and segmentation views are offered, and developers can collect valuable training data without putting hardware, cash or humans on the line.
Microsoft has built the system to be compatible with variety of software and hardware, and preloaded it with some commonly used robotic models and sensors. It interfaces with common robotic platforms, such as the open source Robotic Operating System used by Fetch Robotics, Bosch and others, and the open-source MAVLink communications protocol. Powered by Linux and Windows, the company says its architecture can also be easily extended to accommodate new types of autonomous vehicles, hardware and software.
If you're interested in diving on in, you can head over to Github to check it out.
Source: Microsoft
