We're about to put smart robots in space. A new system in development at the University of Southampton in the UK could allow spacecraft to go beyond simply obeying commands, instead developing and reacting to knowledge about their missions.

A team lead by Professor Sandor Veres has created a testbed for prototypes of the satellites, who learn via natural language processing (NLP) software. They glide across a perfectly level, glass-topped table surrounded by observation cameras and light-blocking curtains. The future space robots perform an oddly beautiful mechanical dance. They glide on precision bearings, their attitude adjusted by eight large fans, guided by inertia sensors and positioning cameras. This is the closest environment they'll get to the weightlessness of space - for now.

The earthbound satellites don't respond to individual commands from a human controller. Instead, they access documents written in NLP. By reading and incorporating these documents into their on-board programs, the satellites gain new analytical and fault-detection capabilities. It isn't quite artificial intelligence, but it isn't a computer just running through a programmed sequence either.


If you're wondering what NLP is, it's a system that allows rules for machine behavior to be written in "natural" sentences. The sentences are natural compared to, say, a few lines of Perl code, but they still require a lot of technical knowledge and precise definitions. You can find an example of an NLP called sEnglish (used with Professor Veres' robot satellites) at sysbrain.org.

While current research is focused on the satellite application, this kind of machine learning could be a major step toward autonomous vehicles on Earth. Someday your self-driving car might rely on sEnglish documents to learn and adapt to new routes and conditions. In fact, sometimes they cover the glass table with fake houses and test their Autonomous Ground Vehicle on it. As Professor Veres explains:


"We have invented sysbrain to control intelligent machines. Sysbrain is a special breed of software agents with unique features such as natural language programming to create them, human-like reasoning, and most importantly they can read special English language documents in ‘system English' or ‘sEnglish'. Human authors of sEnglish documents can put them on the web as publications and sysbrain can read them to enhance their physical and problem solving skills. This allows engineers to write technical papers directly for sysbrain that control the machines."

It's also pretty cool that the robot satellites seem to be built mostly with off-the-shelf electronics.

Source: University of Southampton