THE Cyclopses, according to mythology, were a race of bad-tempered and rather stupid one-eyed giants. Not, perhaps, a great portend for a new generation of robots. But Andrew Davison, a computer scientist at Imperial College, London, thinks one eye is enough for a robot, provided its brain can think fast enough. For a robot to work autonomously, it has to understand its environment. Stereoscopic vision, integrating the images from two "eyes" looking at the same thing from different angles, is one approach, but it involves a lot of complicated computer processing. The preferred method these days, therefore, is SLAM (Simultaneous Localisation And Mapping), which uses sensors such as laser-based range finders that "see" by bouncing beams of light off their surroundings and timing the return.
Dr Davison, however, wants to replace the range finders, which are expensive and fiddly, with a digital camera, which is small, cheap and well understood. With this in mind, he is developing ways to use a single, moving video camera to create continually updated 3D maps that can guide even the most hyperactive of robots on its explorations. His technique involves collecting and integrating images taken from different angles as the camera goes on its travels. The trick is to manage to do this in real time, at frame rates of 100 to 1,000 per second.
The shape of the world pops out easily from laser data because they represent a direct contour map of the surrounding area. A camera captures this geometry indirectly and so needs more (and smarter) computation if it is to generate something good enough for a self-directing robot. The answer is a form of triangulation, tracking features such as points and edges from one frame to the next. With enough measurements of the same set of features from different viewpoints it is possible, if you have a fast enough computer program, to estimate their positions and thus, by inference, the location of the moving camera.
Developing such a program is no mean feat. In the milliseconds between successive frames, relevant information from each fresh image must be extracted and fused with the current map to produce an updated version. The higher the frame-rate, the less time there is to do this work. Rather than throwing more computing power at the problem, though, Dr Davison is using standard processors and concentrating on making his programs super-efficient by analysing the bottlenecks within them and devising ways to cut the number of computational steps. As a result, he and his colleagues have recently been able to show this new form of SLAM working at 200 frames a second on a camera tossed from hand to hand, using just a laptop computer to run the program.
Some well-known games firms are exploring possible uses of the technology, for example to generate a 3D map of a player's room so that it can be incorporated into the game. Dr Davison is also talking to a European company interested in making smart, self-guided vacuum cleaners. If, together, they can create an affordable, dust-fighting robot that can see where it is going (and won't throw boulders at its rivals), that would put the one-eyed myth to rest.