Actually the course also had some mountainous terrain, and yes at least one vehicle went over the edge. I was only partially kidding. On the sonar, I was actually recommending lasers for that task. I realize some use sonar, but I feel lasers are better suited, due to the resistance to added noise. MIT's car can successfully track the lines in the road. So that milestone would be easy to add on to. MIT's car had the bad habit of taking *every* off ramp. The problem isn't so much about being able to read the lane markings, but being able to realize that the lane markings are changing, such as an upcoming off-ramp, or have disappeared entirely. Hence it is better to have a system that can track the road edges, and/or traffic on either side and only use the painted stripes as a general guideline . I think that a system could be successfully built to handle the piano falling off the moving truck. The DOD doesn't care about that. They want to drive heavily armored <insert doomsday vehicle of choice here> over terrain of varying complexity to reach the objective target and take out the bad guys while ensuring it doesn't contribute to "friendly fire". You can add my thousands of hours of thinking about and designing autonomous vehicles to your thousands of hours and the millions of hours of other people. ;') It's a fun challenge. I think it's doable with the PC hardware and technology we currently have. I just haven't had the time to build a prototype to start training. I disagree that it would be easier at night. At night you have to add all the moving lights that may affect sensors in unpredictable ways. I think you'd just be trading on set of noise for a different set of noise, but it does bring up the point that the system needs to be able to transition from one to the other smoothly.

-----Original Message----- From: David Nicol

I realize you are kidding, but I spent thousands of hours in the early nineties thinking about automating driving in cities. My conclusions were:

• it would easier to start with night driving since there are more

obvious indicators and less visual noise, then generalize to daytime.

• collision avoidance would be better done with sonar, as

there is meaningful doppler information available in addition to distance.

• Curb mapping would be handled with sonar too, but visuals are

needed for identifying the lines in the road.

a far cry from the intelligent skateboards in Snow Crash, but essentially dealing with the same problem domain.

IIRC there were successful demonstrations of intelligent cruise controls that kept a distance from the car ahead rather than simply tracking speed, and very intelligent cruise controls that could entirely take over normal highway driving.

Exceptional situations, such as what to do when a piano falls off a flatbed in front of you, are of course unpredictable and will prevent commercialization of the VICC: The potential liabilities are infinite.

I understood the DOD test to be across some rough desert rather than on roads, so the biggest problem is smart obstacle avoidance. Sort of like the science fair robot mazes, scaled up, and not on a smooth floor. Can't have your robot falling off a cliff, unless of course it's built to take falling off a cliff -- like those huge self-propelled beach balls they were talking about deploying on mars.

On Sun, 6 Feb 2005 19:37:28 -0800 (PST), Jack [email protected] wrote:

...

Of course you'd need to add mundane things like crash avoidance, lane changing, turning. All part of the knowledge base based on system inputs. Piece of cake. I could do this in six months with a team of 10 coders. Also, I'm using very cheap labor, probably Rusisans.

Brian Densmore