3 Bots Win Pentagon's Robotic Rally

An anonymous reader writes "We've got a winner in the Pentagon's $3.5 million all-robot street rally, the Urban Challenge. Three, actually. Wired reports that 'bots from Stanford, Virginia Tech, and Carnegie Mellon all completed the course within the six-hour time limit. The robo-cars had to complete different missions taking varying times, so the flesh-and-blood judges will take a day to figure out who takes home first prize."

Comments

Actually, 4 potential

[ZaMoose]

Ben Franklin Racing (a collaboration between UPenn, Lehigh and Lockheed Martin) also finished within the 6 hour time limit.

The judging will certainly be interesting.

We WILL have androids in 20 years

[mangu]

Then it became clear that the algorithms didn't scale well with the complexity of the environment (to put it nicely) and Artificial Intelligence became a somewhat disappointing field for the general public at least.

That's the problem with hype. They have cried "wolf" too many times. It was the same thing at the end of the 19th century, when people were researching flight. Steam engines were too heavy for their power, airplanes had to wait until engines became powerful enough. There were many people, among them some respectable scientists, that wrote articles "proving" that heavier than air flight was impossible.

At this point, computers are too expensive and consume too much power to be practical for anything that involves "human-like" intelligence. But we are making progress, at least we do have unbeatable chess-playing computers, a feat that not so long ago many people considered impossible. Of course, computers do not follow the exact path of reasoning that humans do when playing chess, but they are unbeatable anyhow. Airplanes do not flap wings either, but they fly faster and higher than any bird.

Unless Moore's law ceases to function, we can expect desktop computers with a complexity comparable to that of a human brain in twenty years or so. Given the hardware, it's only reasonable that someone will invent a way to make a computer emulate a human brain in its full power, just like people invented machines capable of flying when they got engines with enough power.

Re:We WILL have androids in 20 years

[inca34]

This has little to do with Moore's and a lot to do with the fact that sensors do not follow Moore's law. We were using the same sensor technology as was available 15 years ago with marginal to no improvement on quality or capability.

The software side of this DARPA Urban Challenge should consist of no more than an enormous, but straightforward, state machine that contains all the logic for traffic decisions. Plug that into a simulator and you've got the main software part done.

The problem everyone had out in the field for the qualifiers (I was on one of the teams) was perception. How do you know what you see is an obstacle? And how do you deal with false positives, and more importantly, false negatives? Some people believe in cross referencing sensor data, which is called sensor fusion. It is difficult, to say the least, to characterize every possible obstacle that ought to be considered a true obstacle if it lies on your vehicles path, let alone have a 10^-6 failure rate for improper detection.

The highway lane following has been solved since the 70s, check out R.E. Fentons work on Automated Highways in Transportation Science (1970). We had some "recent" developments in the early 1990s where we got some autonomous vehicles to do the autobahn at 100mph with more modern sensors and vehicles, but really didn't improve that much because the sensors aren't there yet.

Your sensor choice goes something like this:
$75k for a Velodyne 3D laser system
$5k for the SICK 2D (planar) lasers
~$25k for stereo vision cameras (per set)
~$1k for radar
$75K for the Applanix integrated GPS and IMU

The Velodyne is a spinning set of 64 lasers, with 64 photodiodes. Each manually placed so that the photodiodes are aimed precisely where the lasers are pointed. The entire head of the unit spins at ~2Hz and generates 1 million points per second. Most of the teams that bought one mounted it on top of their vehicle. This sensor is great if you have infinite processing power available to crunch the data and turn it into cost maps. It however has some serious problems: it's very expensive, it's not mass manufacturable, the point data for a rock and a shrub are indistinguishable (a weakness of all lasers), some obstacles we're interested in absorb laser or reflect it away from the photodiodes, it has too much information, and it has moving parts.

The SICK 2D planar lasers have more or less the same problems, except there's less data to crunch of course. These lasers also have moving parts internally, which spin a mirror at maybe ~20Hz to get distance data over a 2D plane. Same issues as the Velodyne, except it's manufacturable (has been for 15 years now).

Stereo vision is really hard to do right. When you have roughly a year to develop the platform and the algorithms, I don't expect much, and I didn't see much. This may be the answer in the future for passive detection, but I don't see it working at the moment.

RADAR is the right sensor for this type of work. It gives you distance and speed. If you're clever it also gives you the "cost" of a particular object. Radar is how you can tell the difference between a shrub and rock, or a car and a plastic fence. The real cost in the RADAR is not the sensor, but the $100k guy who knows RADAR well enough to set it up right and get good data out of it.

The Applanix GPS and IMU with 200k RPM laser gyros are not manufacturable and not practical for autonomous vehicles because of the cost. Perhaps the MEMs solutions will catch up and make IMUs cheaper, but in the mean time we're stuck with these systems if you care about your position.

That's my take on it. Improve the sensors and we'll get autonomous vehicles. Buy another Cray, strap on a generator and a multi-ton air conditioner is not the solution. We need to reliably and cheaply generate cost maps that are relevant to the vehicle that's being automated. Once that's been done reliably, we will have autonomous vehicles. Cheers.