Robots don't see the world like humans do, at least not yet. They don't
recognize discrete objects and have little common sense—that it's better
to drive over a bush than a rock, for instance. Instead of seeing, today's
robots measure. They use a variety of sensors—cameras, laser range
finders, radar—to gauge the shape, slope, and smoothness of the terrain
ahead. They then use these data to figure out how to stay on the road and avoid
obstacles. At least that's the idea, but the DARPA Grand Challenge showed it's
a lot harder than it sounds. Here, see a slide show of the major measuring
techniques used by the various teams.—Jason Spingarn-Koff
|
|
Laser Sensing
Laser scanners, commonly known as lidar (for light detection
and ranging), were the most common sensor for Grand Challenge
robots. A beam of light bounces off a spinning mirror and sweeps the terrain
ahead. By measuring the time it takes for the beam to return, the sensor can
calculate the distance to objects. Line by line, the robot builds a simple 3-D
model of the terrain in front of the vehicle. Many teams used several laser
scanners to increase the amount of detail. Team DAD even built a custom sensor
with 64 spinning lasers. But laser scanners have limitations: They see only a
narrow slice of the world and have a relatively short range (60 to 150 feet).
The beams also can't detect colors and may bounce off shiny surfaces, so it's
hard to spot certain hazards (such as a body of water). They're also bad for
stealthy applications, such as those preferred by the military, because they
emit light.
|
|
|
Video Cameras
Video cameras see fast and far—all the way to the horizon. They can
measure the texture and color of the ground, helping robots understand where
it's safe to drive and alerting them to dangers. And since cameras don't emit
light, they are well suited for stealthy operations. But there's a serious
drawback: it's difficult to use just one camera to figure out the size and
distance of objects ahead. Recent research by Andrew Ng at Stanford suggests
this is possible in some circumstances—by analyzing road edges,
texture, and converging lines—but this needs more testing and wasn't used
in the Grand Challenge competition. Video cameras also have limited use at
night or in dust storms or other bad weather, and they can be blinded by bright
light such as the setting sun.
|
|
|
Adaptive Vision
The Stanford team invented a technique called "adaptive vision," which allows
its robot, Stanley, to see farther and drive faster, even as the road changes
over different types of terrain. Here's how it works: Stanley uses its laser
range finders to locate a smooth patch of ground ahead and samples the color
and texture of this patch by scanning the video image. It then looks for this
color and texture in the rest of the video image. If a smooth "road" extends
about 130 feet toward the horizon, Stanley knows it can speed up. If the road
suddenly changes, the robot slows down until it figures out where it is safe to
drive.
|
|
|
Stereo Vision
Inspired by the way humans see, some teams used two video cameras in a
technique called stereo vision. Princeton, TerraMax, and Team DAD (in the first
Grand Challenge, in 2004) all tried the technique, with varying degrees of
success. In the technique, two cameras are mounted side by side, and software
measures the slight shift between the two incoming images. (You can try this
yourself by holding a finger at arm's length and looking at it only with your
right eye, then only with your left eye; your finger will slightly "jump.") The
shifts are then compiled into a "difference map," which crudely shows the
distance of approaching objects (represented here by different colors). The
problem is that objects in the far distance look largely the same to both
cameras, limiting the accuracy where it matters most: for high-speed
driving.
|
|
|
Radar
These sensors use radar (from radio detection and
ranging) to send out radio waves to a target and measure the return echo.
They can see far into the distance, even through dust clouds and bad weather
(which easily foil cameras and may degrade lidar). The downside is that radar
beams are not as precise as lasers; objects that aren't really obstacles often
appear, confusing the robot. Carnegie Mellon's Red Team used radar as a second
sensor, complementing lidar to spot large objects far ahead. Carnegie Mellon
researchers are developing a more sensitive unit. Called solid-state millimeter
wave radar, this could be a potent sensor on future robots.
|
|
|
Cost Map
Many Grand Challenge robots used this software technique to find the road and
avoid obstacles. First the program compiles data from any variety of sensors
(laser scanners, cameras, radar) and builds a map of the terrain ahead. By
finding smooth areas and pinpointing obstacles, the program can divide the
world into areas for driving that are good ("low cost") and bad ("high cost").
In this image from Stanford, green represents laser measurements, while black
represents unknown grid cells or no measurements. Red signifies grid cells that
are judged to be not drivable (obstacles), while white signifies grid cells judged to
be drivable (road).
|
|
|
Path Planning
This software program allows the robot to find the best path from point to
point, even as it avoids obstacles. Some robots use a searching algorithm, a
sort of mathematical recipe, to figure out every possible path, then compare
each to the cost map to find which path is best. They must also factor in how
fast to go and which paths are physically impossible (e.g., sharp turns could
cause a rollover). As they're driving, the robots must constantly update their
paths to stay on the racecourse and steer toward the finish.
|
|
|