Stanley and the Conquest of the DARPA Challenge

geekboy_x writes "Wired has a great in-depth piece on the Stanford team that won the $2 million DARPA prize. If you remember last year's disaster - with most vehicles falling off the road in the first kilometer or so - this victory becomes all the more amazing. The fact that the Stanford team used a 'tailgating' strategy is the best surprise in the article."

Comments

Team Leaders

[Kuxman]

Also interesting to note is the fact that the major leaders of the Stanford team came from the Carnegie Mellon AI department 2-3 years ago.

Re:Team Leaders

[maggard]

This could as easily imply that, in order to succeed, these folks had to get out of Carnegie Mellon AI and go to Stanford .

I've no inside knowledge, but from the article it appears CMU was locked into the-same-just-more/bigger/faster strategy and the team that decamped to Stanford came up with some innovative real-time confidence-based sensor interpretation systems. It may well be that at CMU they wouldn't have been supported in this whereas at Stamford, without the established regime at CMU, they were free to do so...

Why not flying cars, then?

[Radres]

FTA: "He liked to point out that planes had been flying themselves since the 1970s. The public was clearly willing to accept being flown by autopilot, but nobody had tried the same on the ground."

Just give us our flying cars then already, damnit!

Re:Why not flying cars, then?

[Kuxman]

Actually, the Boing 777 does land/take off automatically. I think this also holds true for the Airbus 300s (Correct me if I'm wrong)

From "Ask Captain Lin":

"On the Boeing 777, the autopilot can be selected on at 200 feet above ground level after take off. Most of the time, the pilot would make use of the autopilot on the climb because it eases the workload of the crew especially during an emergency. Sometimes, a pilot may elect to fly manually during the climb just to get his hands on the control column or to maintain his proficiency because during a flight test, one of the exercise calls for flying without the aid of autopilot. Otherwise, the autopilot is engaged throughout most of the flight. It is smoother, more economical and safer with the autopilot on. In fact, in really bad weather with very limited visibility, the autopilot even lands the aircraft by itself. The pilot only resumes control of the aircraft after it has safely landed on the Runway."

Re:Why not flying cars, then?

[Zocalo]

My cousin is a qualified pilot on several of the bigger passenger jets and yes, it is entirely possible for a crew to do nothing but board the plane, taxi to the runway and then let the autopilot handle the entire flight, including the takeoff and landing. The normal mode of operation however is to clear the airport on manual, activate the autopilot until in the approach at the destination and then make a judgement call about letting the autopilot land the plane at the destination based on the conditions at hand. There are also exceptions about if one or more of the autopilots malfunctions (there are apparently three on the bigger jets, I'm not sure about the smaller ones). Technically one functional autopilot is enough to handle the entire flight, but the regulations of my cousin's employer prohibit non-manual landings with just one faulty autopilot, and with two faulty units all flight operations must be fully on manual. They do however have to complete a mandatory amount of manual take-offs, landings and flight hours each year to remain qualified, in addition to the numerous medical, physical and flight examinations you would expect. Other airlines do vary their individual guidelines and proceedures of course, but not by too much.

Re:Why not flying cars, then?

[jackb_guppy]

And in fact all AIR based accidents end up on the ground or below.

The most interesting aspect of the article...

[NeutronCowboy]

... is that the CMU team relied heavily on extensive pre-analysis of the environment, and failed (at least in the sense that it didn't come in first). Stanford instead relied on a probability analysis of the incoming data, along with multiple technologies for different goals (lasers for short range data, video for long range data).

It seems that the DARPA grand challenge not only showed off the first realistically autonomous vehicles, but also laid to rest the idea that expert systems were the way forward. The way forward instead is self-teaching computers. Hooray for self-teaching AI overlords!

The surprising thing is the good vision system

[Animats]

As one of the team leaders of another Grand Challenge team, I'm enormously impressed with the Stanford work. The basic idea is that the LIDARs profile the road ahead out to 20m or so, and the vision system decides whether the road further out is "like" the near road. That vision system was a huge breakthrough. It was obvious that such a system would be a big win, but making it work reliably was impressive. I didn't think that was possible at the current state of the art. I look forward to seeing a more detailed paper on how it was done. A good hint is in this paper on texture comparison.

I was never that impressed with the CMU approach. All that manual preplanning was an obvious dead end. And the giant mechanically stablized gimbal was just too clunky. It didn't help them in 2004, when they hit an obstacle placed by DARPA, and it didn't help them in 2005, when DARPA moved the racecourse from California to Nevada to prevent preplanning. The Air Force colonel in charge for 2005 said preplanning wouldn't work, and he meant it.

Computer vision of the natural world is finally about to take off, after three decades of frustration. It's probably possible to do much of the early vision processing in a current-generation GPU, which may make it affordable. Look for new apps that connect to cameras and pick out items of interest. Read that paper linked above.

The part of TFA that floored me

[sikandril]

was when Thun explained how the vehicle was taught to drive by following a human driver and adapting its algorithms according to his behavior, gaining much better results than "force feeding" massive amounts of data artificially.

This has immediate implications not only for robotic cars - what if we took a human and strapped some positional sensors, voice recording, etc. and made a humanoid robot follow him throughout the day?

I mean how varied are our lives after all? Given the right processing power and sensors, the results could be interesting...

Again, a great achievement for a 'bottom up' approach to artificial intelligence

No tailgating. Wired has it wrong.

[Animats]

That's actually not true. There was no "tailgating". During the Grand Challenge, no vehicle was allowed to approach another while both vehicles were active. DARPA had the ability to remotely pause any vehicle. When vehicles got anywhere near each other, the trailing vehicle was paused to maintain separation. If the trailing vehicle was clearly faster, a pass was scheduled. All passing took place with one vehicle stationary and at a wide place in the road. Wired has this wrong.

Static problem

[kurtkilgor]

As a participant of another DARPA team (Cornell -- our site is down), I am skeptical as to whether the winners of the challenge would be able to drive in a real world environment. In many ways the Grand Challenge was a toy problem, but this is not usually emphasized because they want to make it seem more dramatic.

First of all, no other moving objects on the course. When a vehicle was about to pass another, the one in front was paused so that the passing vehicle could overtake it. At no time did the vehicles have to deal with changing conditions.

Secondly, to my knowledge, there were no obstacles (which were promised) on the course. If someone knows differently, I'd like to hear about it. So we don't know to what extent obstacle avoidance is effective on those vehicles.

Thirdly, daylight and clear weather is one thing, but nighttime, rain, snow, etc. would significantly degrade the data.

Essentially the problem that the current vehicles solved was this:
Given a set of waypoints and a "corridor" outside which you will never have to go (so far the problem can be solved only by 10cm-accuracy DGPS), use your other sensors to avoid obstacles by moving left or right within the corridor.

Not very much like real world driving at all. And I'm not saying Stanford, CMU and the others didn't accomplish something big -- I'm just saying it's not what the Wired piece makes it out to be.