Giving mobile robots the ability to remember sites they have previously visited is crucial if they are to build accurate maps. Now robots built in the UK are using a human-like trick to build maps of journeys up to 1000 kilometres long 鈥� the largest robot maps yet produced.
Robots equipped with laser scanners or other ways of viewing the immediate environment have been able to chart their progress for years by plotting orientation and distance travelled. But those robots can all too easily get confused and fail to recognise a place they have been before, potentially producing hugely distorted maps in the process.
Now researchers have come up with a solution that is similar to the way people view the world and means their robots can more accurately recognise places they have been before 鈥� even when objects have moved or are approached from a new angle.
Advertisement
Spotting changes
Recognising a place you have visited before is not as straightforward as it might seem, and from the University of Oxford told New 杏吧原创.
For a robot with no grasp of what the objects around it are, the world is a chaotic and dynamic place. A robot may not return to a point for several hours, or even several days, so it needs to be able to ignore minor changes such as cars parking or driving away. However, it must also avoid accepting so much change that it falsely recognises a new location.
The Oxford group鈥檚 tackles those problems by having a robot assign a visual 鈥渧ocabulary鈥� of up to a thousand individual 鈥渨ords鈥� for each scene, every two seconds.
Word bags
The 鈥渨ords鈥� describe particular objects in a scene, for example a bicycle seat, and the software learns to link words that occur together into groups that are given words of their own. For example, the word 鈥渂icycle seat鈥� is almost always found associated with the words 鈥渂icycle wheel鈥� and 鈥渂icycle chain鈥�, so they linked together in a so-called 鈥渂ag of words鈥� 鈥� 鈥渂icycle鈥�.
That means when the robot revisits a scene that now lacks, say, a bicycle, it notes a single change rather than the disappearance of many smaller features. That prevents too much significance being attached to the bike鈥檚 disappearance and means the robot is more likely to recognise the scene as familiar, says Newman.
This hierarchical approach reduces the chance of falsely recognising new scenes too, he says. Typically, a navigating robot would confuse two similar looking brick walls that each contained a window. But the way FabMap lumps together similar objects prevents that.
Instead of being swayed by 鈥渞ecognising鈥� several hundred similarities in row upon row of near-identical bricks, the software counts them only as a single feature. More reliable points of comparison, like the height of the windows, carry more weight.
Bot about town
The researchers set loose a mobile robot equipped with FabMap, cameras and a laser scanner in Oxford city centre. Even though the environment was subtly changing all the time, the robot was able to recognise familiar terrain and adjust any errors in its map, while only once giving a false positive out of an image set that included 10,000 photographs.
In ongoing work, the team has strapped the kit into a car in order to map 1000 kilometres of Oxfordshire countryside and roads 鈥� the largest scale maps made by any robot to date.
, an applied researcher at Microsoft, thinks the work is impressive.
鈥淎 handful of researchers have been active in the area of vision-based mapping for [most] of the last decade, and [this] work stands out as a sound marriage of theoretical principles and practical engineering,鈥� he says.
He is particularly impressed with the robot鈥檚 ability to instantly recognise a familiar scene even after accruing a huge amount of visual data on its travels 鈥淣ot more than five years ago we were focused on building visual maps of just one or two rooms in a building,鈥� he says.
