Stanford Robot Car Capable of Slide Parking

kkleiner writes "Stanford's Junior, the robot car that took second place at DARPA's Grand Challenge in 2007, has learned how to perform a tire-squealing 180-degree spin into a skin-tight parking space. Similar to a James Bond action scene, the maneuver is impressive and would be extremely difficult for a human to pull off. We won't be handing the keys over to robot cars anytime soon, but Stanford shows us that at least for some driving tasks robot cars can already meet or even exceed human ability."

Comments

Three Points

[XPulga]

1. The proper reference is the Blues Brothers movie, not James Bond.

2. Parking like this is stupid and wears down the tires unevenly and too fast.

3. Uneven pavement, potholes, wet pavement, oil puddle: pick your disaster.

Re:Three Points

[JWSmythe]

    On #2, I ripped a front tire from it's rim in a Crown Vic once doing that. Well, it was actually a 30mph J-turn, but still the same general movement. Low profile tires help reduce that. In my case, a fun J-turn on a closed course (performance driving training pad, after hours), turned into an exercise of changing a tire when it was 95 degrees out with the sun beating down on you.

    I wouldn't worry about the uneven treadwear. I'd worry more about the stress it's putting on the suspension and steering.

    At the speed and distance they executed that maneuver, and since the driver went past the space already, they would have been able to evaluate the situation. I don't know how well their computer is going to evaluate the road conditions. Did it hear the loose gravel tapping under the car when it drove through the first time. Us humans take all kinds of clues from our environment to make our decisions.

    In performance driving (closed track, of course), I'd have to say a human will still be far superior to a computer. At one course, because I was a "newbie" to their track, I had an instructor with me (their rules). I ran the track hard for 1/2 hour, and got used to how the whole course handled. I parked for about an our while other sets of folks ran. When I went back out, since my tires were now cold, when I hit the first turn (a tight 270), the back end kicked out. I drifted the turn, rather than spinning or sliding off the track, and took it easy for the next couple laps until the tires warmed up again. When we got back to the pits the instructor said to me, "I didn't think you'd make that turn. Most of the people out here would have lost it."

    Us humans learn a lot from practice. I knew immediately what was happening from the way the car responded. It could have been a bunch of other reasons, but I knew my car, and I already knew how it should respond on the track. It wasn't because I'm an amazing driver, or my car is the best in the world, but through experience I know what those little feelings mean. I kind of cheated. When I was a kid, I'd take my 4,800 pound car (that cost a whopping $300 used) with an amazingly sloppy suspension on grassy roads (laid out but never developed) after it rained and would intentionally spin it so I could practice recovering. At first, I'd end up off the road. With practice, I learned to not stop sideways, but to actually bring it under control and continue driving. But hey, I was 16, and my friends loved to ride along while I practiced.

    I guess with all that said, maybe computers will be able to outperform the average driver. I've seen enough people do enough stupid things because they didn't know what else to do, and never went to learn more about how to make their tons of steel drive down the road safely. I wish more drivers would go through a fraction of the proper training that I've done. Most people don't care. "I start the car. I point it to where I want to go. I get there." As long as the computers don't do as badly as the recent Volvo collision avoidance braking demonstration ("truck? What truck?"), I'll be impressed. :)

Re:This is awesome but...

[shutdown -p now]

IF you read TFA (a novel concept, I know!), it has a longer video which demos several different algorithms which fail variously; and then, ultimately, a final run which combines all of them to succeed. They claim that it is this smoothless combination is what is the real innovation here.

Stanford hasn't heard of gymkhana, apparently.

[SuperBanana]

Similar to a James Bond action scene, the maneuver is impressive and would be extremely difficult for a human to pull off.

Bullshit. Yeah, he's one of the best- but he's doing that in a 500hp AWD car, not a 100HP FWD diesel station wagon, at speeds several times higher than what Stanford was doing. Call me when they can do what he does.

It's also extremely difficult for a human to pull off crochet if they haven't been taught how. Or to shoot a rifle and hit a target a mile away. Or fly fighter jets in formation feet apart. Yet we do that. The question is: how hard is it to train someone, and how consistently can they do it, and how much effort did it take to get the computer to do it?

The answer to the first part: Top Gear did a show segment where they had Russ Swift teach a bunch of people off the street how to do it. If I recall, they were grandmothers. They were going for a larger area, but come on- they were octogenarians.

The answer to the second part:

Apparently Stanford hasn't heard of rallying or gymkhana. Tens of thousands of people do stuff way, way more impressive than what Stanford is demonstrating- at much higher speeds in much more powerful cars. It's not hard, and the Stanford guys are grossly overexaggerating the complexity of the problem to model, as well. The whole point is that you use the car's momentum and lock wheels to make it slide predictably. Practice makes perfect for timing and aim (in the case of Top Gear, they practiced with inflatable boxes that were harmless to the cars.)

And, how many tries do you suppose it took the Stanford team to get it right?

Re:This is awesome but...

[shutdown -p now]

I think that is really cool AI. But I don't know if failing a few times at a driving maneuver is really going to work for me as a passenger though.

To reiterate: the failures (on the video) were with a different AI (which they claim is the "conventional" algorithms used today). The success was with their AI, which is a combination of those techniques that, individually, were each failing on its own. There's nothing in the video that implies that their AI fails intermittently at this point.

Of course, I'd imagine that they've spent a lot of time testing & debugging it, and those skid marks could have come from there as well.

Thanks for checking this out

[zkolter]

Hey everyone,
Thanks for checking this out! To answer a few questions that have been asked:

This video actually was indeed shot the first time we put the whole system together. Of course there were other runs, both demonstrating the alternative approaches and before we had everything working properly, that didn't succeed, but the final system was pretty reliable as autonomous driving goes. That said, we'd want to test this quite a bit more before I'd be willing to lie down where those cones are, and a big issue here is that the maneuver does shred through tires pretty quickly and is pretty tough on the car in general :-).

Second, I certainly wouldn't argue that what we're doing here rivals the very best human drivers (the claim we're making is just that this is one of the more challenging _autonomous_ maneuvers that has been demonstrated). The best humans are certainly able to drive incredibly impressive stunts, and we only claim to be making progress towards this level of ability. However, it's worth noting that this particular maneuver is probably one that _most_ people would have trouble with (I know I certainly can't do it!).

Let me know if there are any other questions, and I'll do my best to clarify.

Thanks!
Zico