Microsoft wants to build software that teaches drones, robots, and drills how to think

Microsoft Corp. and technology rivals spend a lot of time talking about machine learning. Now Microsoft is talking about something called machine teaching.

No the software maker doesn't plan to send robots into classrooms. In a world where factories and wind farms will increasingly run on autonomous systems, drones will criss-cross cities delivering packages and robots will operate in underground mines, Microsoft wants to make the software that helps mechanical and chemical engineers teach those devices how to behave, where to go and how to maintain safe conditions.

Microsoft last year acquired a company called Bonsai that makes this kind of software, merged it with some work from its research arm — a group of Microsoft researchers wrote a paper on this idea back in 2017 — and  is now expanding a software preview so more potential customers can test it. As the company tries to sell more of its Azure cloud software to industrial companies, it aims to make these kinds of autonomous programs a profitable part of that portfolio. Many consumers will be most familiar with this kind of software as it exists in self-driving cars, but Microsoft plans to leave that part of the market to the Teslas of the world.

Delta Air Lines Inc. is running a project to improve their baggage handling using the technology. Royal Dutch Shell Plc is trying out the software to control drilling equipment, while Schneider Electric SE is seeing how it works with electric heating and cooling controls for buildings, said Mark Hammond,  founder and chief executive officer of Bonsai, who is now a general manager at Microsoft.

A Microsoft partner based near that company's Redmond, Washington, headquarters wants to use it for tractors and Carnegie Mellon University deployed the software as part of a mine-exploration robot that recently won a Defense Advanced Research Projects Agency challenge.  Microsoft has also suggested the software could work well for drones that check power lines and wind turbines and for disaster recovery operations where autonomous devices scout out the situations that may not be safe for human rescuers.

"The industry is fixated on autonomous driving and that's it, but if you look around you in the world, you can find literally hundreds and hundreds and hundreds of scenarios where automation can improve things," said Gurdeep Pall, Microsoft vice president, business AI.  "A lot of these folks who build these systems are mechanical engineers, electrical engineers, etc. They are not AI people. We are bringing AI to these engineers in a way that they can operate."

In May, Microsoft began a limited preview of the software that extended to about 50 customers. The software allows engineers to set up rules and criteria for how autonomous devices should operate, anything from where a robot arm should start, what it should do next and all the different possibilities. Then engineers use simulation software — either from Microsoft or its partners — to set up a series of lessons, a digital curriculum. "It's not randomly exploring, it's exploring in a way that's guided by the teacher," Hammond said. And once you have the curriculum, the system automates the process of teaching and learning, across hundreds or thousands of simulations at the same time. 

Microsoft partner Fresh Consulting is working with several customers to figure out how to program devices and vehicles with Microsoft's tools. One such customer is  industrial equipment rental company United Rentals, and Fresh wants to use Microsoft's product to better control compact track loaders, which need to work in uneven terrain and mud. The software can also be useful in construction and warehouse work. "These are dirty, dangerous and dull jobs, and there's not enough people, said CEO Jeff Dance. 

Microsoft is also partnering with MathWorks Inc., which makes simulation and modeling software used by companies like Toyota and Airbus, to allow its programs to work with Microsoft's. 

Microsoft said its autonomous software approach blends the power of human experience with the ability to adapt to changing situations through a type of AI called reinforcement learning. For example, Shell is using the tools to teach its drills. Shell could program drills the old fashioned way, with a series of rules put in by the human experts, Hammond said. But that would require lots of time reprogramming each drill every time it's used on different terrain. A reinforcement learning system — like those used to teach machines how to play video games better than humans — could learn how to do it alone. But for industrial tasks, reinforcement learning with human knowledge and guidance works better, Hammond said. Without it, systems may come to conclusions that don't make sense in the real world. 

Software for factories, equipment and industrial applications is often very specific and made by companies in those industries rather than large, general purpose software makers like Microsoft. And many of those vendors are also working on systems for increasing autonomous control. Microsoft also wants to sell other products, from cloud services to HoloLens augmented reality goggles to construction and industrial firms. Meanwhile its cloud rival Amazon.com Inc. is trying to leverage expertise in logistics and warehouse automation to sell services to industrial companies, said Nick McQuire, an analyst at market research firm CCS Insight. Amazon and Google are also working on AI learning techniques with robots and on programs that promise to enable engineers without AI expertise to program complicated AI models. 

Rather than try to compete with industrial tech vendors, Microsoft wants to partner with them, Pall said.

"It's a big market, but a very difficult one to target in terms of the complexity and the legacy systems, and a lot of those systems are highly mission critical," McQuire said. "It's going to take some time, but Microsoft is starting to position a lot of its products for it."

(with inputs from HT Correspondent)