factories where it pays off. So if you have a large factory, where the roots are stable, then it’s worth paying the hundred thousand or so it will probably going to cost you to get the robot installed. But in most other places things change
too often, and it’s just not worthwhile bringing in a person to program the robot each time something changes. Besides that the factory owners are very nervous about having to depend on somebody outside who may not be there next year.
So it was a very hard sell. Only dozens of these insect-like advanced robots were sold. Ultimately, I guess, there were hundreds. Most of the companies just went out of business that tried to do this in the Eighties. Then in the late Eighties and early Nineties another kind of robot started appearing in research settings. This robot didn’t use navigation techniques like I’ve just described to find its way around. Instead it was able to map the world around it using sensors in a very general way, and was able to actually navigate by these maps that it built itself. In principal, with the right high-level program, these robots could be put into an entirely new place, and still do the right thing.
We worked on that in the Eighties ourselves, and still into the Nineties. All that was possible with the amount of computer power that we had then–about a million calculations a second–was the ability to build maps like this in two dimensions. But this handles ninety percent of the problems, because if you build a map that’s at the height of the belt-line of the robot, it’ll contain most of the obstacles that it’s going to meet–the main walls and furniture. This will allow you to both plan sensible paths by matching up large areas from one time to the next, to localize yourself from one time to another.
So for instance, you’d have a robot that someone would lead through a path, and it would memorize the maps that happened along that path. Then, next time, when the robot was on it’s own, it would just recall those maps that it memorized during the training, and match them up to the current maps that it was getting–slide one against the other until they lined up the best. Then it would know where it was now compared to when it was trained, and it would know where it should be. But it took just about all the computation it could do with one MIP, and it was almost not possible to do in real time, or just barely possible. By the early 1990’s reasonably inexpensive computers had gotten up to about ten million instructions per second (MIPS), and then the two-dimensional maps got relatively easy to do.
Nowadays you can find a lot of robots cruising research hallways that do their own mapping. They wouldn’t have to be specially installed if they were used to do something practical–except for one problem. With a two-dimensional map there are places where the world gets ambiguous and the robot can be confused. A two-dimensional map is typically made up of cells, dividing the world into a grid of cells, and noting what’s in each cell–sometimes just where it’s empty or occupied, or a probability that it’s occupied. But it only had a few thousand cells, and with such a fuzzy and low resolution picture there are ways for that picture to be wrong, which makes it look like another picture as it were. So the robot can be confused about where it is, or it can miss important things in its surroundings. The chance of that is fairly low, but when a robot’s cruising around for hours, days and weeks, even a low probability eventually bites. The mean time between even the best two-dimensional mapping robots screwing up seems to be about a day–which means its good for a demonstration, but it’s not good for practical use.
These insect-like robots that began to appear in the Eighties can be installed on a company’s premises. These are robots that deliver, clean floors, or act as a security guard. If it works fine for a month, but then at the end of a month it gets into trouble, wanders down the wrong corridor, gets stuck in a corner, or in some cases falls down the stairs, the customer is no longer interested in it (laughter), and it’s out the door. It’s had its one month of testing and it failed. But if the robot manages to do its job for about