Microsoft Research has always worked on some of the most crazy-yet-practical ideas and projects. Most of these projects don’t witness the light of day, but some do – like this one.
There was once a time when developers and programmers used to create very specific conditions to “teach” a computer or machine certain tasks. It’s wasn’t a very flexible way of doing things.
The solution came about as machine learning. The term is quite literal – the computer is taught, much like a child, what certain things are and how they work. The technology has been used extensively in a multitude of products and services by various companies.
Bing and Google use it to improve image search results, while Tesla, Uber, Ford, and others, use it to make self-driving cars safer.
The machine can learn what something is, and how it functions. But there is a problem: to teach the machine, designers and developers need to feed a lot of data into these systems.
It’s a repetitive task – like learning always is – but machines need a lot more data to perceive something than a human child does.
Aerial Informatics and Robotics Platform
Machines can get quite clever, but the cleverness can be at the cost of various risks. Teaching a car how to drive itself or teaching a drone how to fly itself require data from the real world. It’s one of the reasons Google, Tesla, etc. collect data from real-world cars driving on real roads and highways.
This data is fed into the system, where the machine can learn from it. That’s where the trouble is.
Spending thousands of hours flying a drone in the real world is not only a tedious task but also a risky one.
So, Microsoft made a realistic simulation of the real world and made it open source, because that’s the obvious solution.
That’s what the Aerial Informatics and Robotics Platform is. It provides a realistic simulation and accompanying tools to help designers and developers generate the insane amounts of data they require to train their automated systems.
Simulations like this have existed for a couple of years, but Microsoft claims that its platform is far more advanced. It’s leveraging the latest advancements in graphics technology and incorporates realistic physics and perception to create an accurate simulation.
In layman terms: shadows in the simulation look just like shadows in the real world. Therefore the data generated in the simulation can be utilized to train drones for the real world.
That’s a very simplistic explanation, but it’s accurate.
“You can do a lot of experiments, and even if those experiments fail they have very little cost in real life,” says Ashish Kapoor, the senior Microsoft researcher in charge of the project.
Microsoft’s solution doesn’t simply reduce the cost, but it also makes the entire ordeal less tedious.
The platform allows developers and designers to quickly make changes to their algorithms and test it in the simulation. That’s much faster than doing the same in the real world and allows for rapid prototyping.
Microsoft has made available a beta version of the tool today. It’s free for anyone to try and use, and is available on GitHub.
You can read more about the project over here.