Robots that use educated guesswork to build maps of their surroundings are being tested by US researchers. The approach could let them navigate more easily through complex environments such as unfamiliar buildings, the researchers claim.
Navigation is one of the biggest challenges faced by mobile robots. One popular technique, dubbed SLAM (simultaneous localisation and mapping), involves having a robot build a map of the local area, whilst also tracking its position (see Uncharted territory).
While humans find it easy to create 鈥渕ental maps鈥� in this way, it is difficult and time consuming for a robot to perform the same task.
Advertisement
Robots typically use laser scanners and odometers to measure distances for mapping. To speed up this process, and to make it more accurate, researchers have previously tried using different algorithms, or set teams of robots to explore an area together.
Now, George Lee and colleagues at Purdue University, US, have come up with an altogether different approach. They have developed an algorithm that uses information already collected to 鈥済uess鈥� what comes next.
鈥淲e realised that, because you are building up a map as you go along, you can use it like a database to predict the environment in unknown areas,鈥� Lee told 快猫短视频. 鈥淥nce you have that prediction, you can either save time and not look, or explore anyway and get a more accurate map.鈥�
Into the unknown
The team鈥檚 algorithm identifies unexplored regions, known as 鈥渇rontier cells鈥�, adjacent to areas that have already been mapped. It then uses the pattern of corners at the edge of this cell to search for similar patterns that have already been mapped. If a match is found, the algorithm uses the existing map to make a prediction of the contents of the frontier cell.
Each prediction has a 鈥渃onfidence score鈥� attached to it. Areas with a high score can be left unexplored to save time, while predictions with lower confidence scores may need to be mapped properly.
The algorithm was initially tested using simulated robots, placed inside virtual mazes and office environments. The simulated robots were able to navigate successfully while exploring 33% less of their environment.
Real-life tests where then carried out with small robots inside an office building at the university. The real robots also saved time and experienced fewer mapping errors, thanks to combined predictions with measurements made using onboard instruments (see image, right). Building maps using measurements alone is less accurate because instruments are prone to errors.
Sharing data
Lee and colleagues plan to extend the method to multiple robots. 鈥淵ou could have two robots building their own maps,鈥� he says, 鈥渨hich then share them when they meet.鈥� This will allow a robot to make predictions based on data collected by its teammate.
But the method does have limitations, Lee says: 鈥淚t works well in indoor environments, but wouldn鈥檛 be very good in less-repetitive outdoors environments.鈥�
鈥淭his approach makes sense,鈥� says Andrew Davison, SLAM researcher at Imperial College London, UK. 鈥淩ather than getting all the detailed information, they鈥檙e learning static, repeating features of the environment.鈥�
鈥淚n a way it鈥檚 getting close to the way humans build a map,鈥� he adds. 鈥淵ou store background information and can understand new areas that are familiar to old ones.鈥�
Davison and colleagues are designing endoscopic surgical instruments with SLAM abilities. Like Lee鈥檚 mobile robots, the instruments could use other information to make predictions about their environment.
鈥淚t is a very hard environment to map because of the large amount of movement and lack of clear features,鈥� Davison says. 鈥淒rawing on pre-operative scans could make it easier.鈥�
Journal reference: IEEE transactions on robotics (vol 23, p 281)
