Robotic Nation Evidence

One of the teams in DARPA's Grand Challenge race used an innovative technique for object avoidance. This technique shows what is possible as more and more cheap processing power becomes available through Moore's Law. Here what Popular Science magazine had to say about the technique in the June 2004 issue:

Digital Auto Drive of Morgan Hill, California, developed an innovative new robot vision system that, team leaders claim, nearly won them the race.

In theory, the team's strategy is simple: Avoid the struggle to combine data from multiple "active" sensing systems like radar and lidar. Create a real-time 3-D map of the course in front of the vehicle with video cameras alone. Choose a safe path through the upcoming terrain, then point the steering wheel down that path. Now do it while flying through the desert at up to 60 mph. The team, led by engineers David Hall and his brother Bruce, mounted two digital video cameras atop their 2003 Toyota Tundra. A processor transforms the images from the cameras into a 35-billion-pixel terrain map, and redraws that map 60 times a second. For each one of these updates, another processor creates up to 100 possible paths through ifie terrain, then rates each path based on how flat it is, how close it is to the DARPA-defined course, and the size of the obstacles on that path, (The team completely avoids the problem of object recognition by treating every bush like a boulder, and avoiding both.) Repeat the process 60 times a second: 35 billion pixels, 100 paths.

By today's standards, that's a big computational load. Let's assume a simple, efficient algorithm that uses only 10 processor operations per pixel. 35 billion pixels means 350 billion operations. 60 times a second means 21 trillion operations per second.

Obviously that is not how they do it -- even the largest supercomputers today are not much faster, and they are immense. NEC's earth simulator can only support 35 trillion floating point operations per second. It draws megawatts of power to do its job. It certainly would not fit in a truck.

But let's say they do want to use an algorithm that consumes 21 trillion operations per second. Today that takes perhaps 10,000 Pentium 4 chips. In 20 years or less, it will take only 10 chips. And this does not consider the fact that a Pentium chip is general purpose. A single chip will be able to handle the load by tuning it to the vision-processing application. By then, there will also be algorithms that can detect the difference between bushes and rocks 60 times per second. There will be algorithms that can recognize endangered turtles and avoid running over them. And so on.

The thing that is holding back computer vision right now is processing power. As more and more processing power becomes available, innovative algorithms will exploit it. It will not be long before cars can see the road better than people can see it, and our cars and trucks will drive themselves far more safely than people drive them today.