Page 157 - Modern Robotics Building Versatile Macines

RADICAL ROBOTICIST 137

SOLVING PROBLEMS: IN THE DRIVER’S SEAT

The work being done at Carnegie Mellon University and elsewhere
on robot driving systems has many potential uses. Eventually, if
systems can be made safe enough, self-driving cars could greatly
improve the nation’s most popular form of transportation—the
automobile. Special freeway lanes may be set aside for automatic
driving, and the computers could optimize traffic into a smooth
flow and maintain fuel-efficient speeds. The human driver could
become a passenger, free to catch up with some work or just enjoy
the scenery.
In the late 1990s, the National Automated Highway System
Research Program, a consortium of government agencies and private
corporations, demonstrated the feasibility of automated cars. A num-
ber of serious obstacles remain. For example, if the carrying capacity
of highways is greatly increased by automation, what happens when
all those extra cars exit from the highway to already congested
surface streets? In general, the interaction between automatic and
manual drivers remains a problem: Tests of recent Navlab vehicles
showed that the computer could react to road conditions much
faster than human drivers but was not as good at predicting what
those crazy human drivers might do!
The most likely practical results of the automatic driving research
will be technologies that assist but do not eliminate the human
driver. For years to come, increasingly sophisticated “cruise controls”
and systems that can warn inattentive human drivers that they are
starting to veer out of their lane or are getting too close to other
traffic will appear.

Looking back from the perspective of 2003, Moravec, in his
paper for the Association of Computing Machinery, “Robots, After
All,” acknowledges that researchers in the 1970s and 1980s under-
estimated greatly the visual processing capacity of the human brain.
During those two decades, even the seemingly powerful computers
in the leading laboratories could only perform about 1 MIPS (mil-
lion instructions per second). Moravec notes that