Robot Parking

by on July 17, 2010 at 7:12 am in Science, Travels | Permalink

In the first DARPA Grand Challenge for driverless vehicles in 2004 not a single team came close to finishing the course.  Later this year a driverless car built by a team from Stanford will race up Pike's Peak at speeds up to 90 mph.  Amazing.  And from the same team, I would pay for a car with this type of automated parking.

1 Slocum July 17, 2010 at 9:46 am

I’ll be impressed when and only when:

1) Robotic vehicles can do this with NO GPS assist or built-in maps of any kind (that is, that they navigate only using artificial vision, radar, etc). GPS is cheating.

2) They can handle unexpected, non-fixed, non-stationary objects along the road — when they can, for example, distinguish between a few cow behind a fence (that can be safely ignored) and a few deer alongside the road that might dart out at any time.

Maybe some of these automated systems will prove useful for driver controlled cars (e.g. GPS-based systems could warn or even slow the vehicle when approaching a curve or t-intersection at excessive speeds), but I really don’t see much use for truly autonomous vehicles until they can recognize and comprehend the environment they’re moving through and the objects in it — which presumes turing-test level AI capabilities that are far beyond the state of the art.

2 david July 17, 2010 at 10:55 am

Why would GPS be cheating? Cars need to be able to plan a route. Likewise we should arm driving robots with radar and thermal imaging and infravision. And mesh-networked sharing of information with nearby cars. And any other kind of imaging and mapping methods engineers think up.

Differentiating cows behind a fence and deer should be easier when a car can ‘see’ deer, cows, and fences using thermal imaging for the ungulates and radar for the fence. Or whatever.

Even modern automobiles need considerable infrastructure to be useful, especially in the presence of other traffic.

3 Calvin July 17, 2010 at 12:24 pm

I believe (though I am not absolutely certain) that GPS was not allowed for the first DARPA challenge, which was one of the reasons the vehicles failed so miserably.

Following GPS waypoints using onboard GPS was absolutely part of the challenge, and the slightest bit of rudimentary research would have told you that.

4 anonymous July 17, 2010 at 12:49 pm

If and when this ever becomes practical, it will be a very short step until insisting on driving your own car will be stigmatized as severely as drinking and driving…

5 jim July 17, 2010 at 3:31 pm

Proximity sensors and automatic breaking would be nice. I’m speaking as somebody who got run down from behind by a car in a parking lot. I was walking to my car and the driver didn’t see me. They only hit the brakes after I smashed their windshield with my head.

The car was going slow, but a car can do a surprising brutal amount of damage even at slow speeds. Not sure exactly what happened since my memory of the event is thankfully non-existent. It ended with me leaving the store and thinking, hey I think I’ll go to my car now. Starts up again when the paramedics arrived. In between … not so much.

6 agnostic July 17, 2010 at 5:30 pm

If it were amazing, then people would be amazed. Cars stopped amazing people decades ago.

7 Miller July 18, 2010 at 12:27 am

Slocum

I doubt anyone cares much whether you’re impressed. And no one is going to refuse to buy an autonomous vehicle that uses GPS on the grounds that GPS is “cheating.” I see no basis for your claim that true autonomous vehicles will require “turing test level AI capabilities.” The DARPA Urban Challenge demonstrated that we can already build vehicles that can drive autonomously in an urban environment, obeying traffic laws, avoiding obstacles, and sharing the road with other vehicles, both robotic and human-operated. Commercially-available street-legal fully-autonomous vehicles are probably at least a decade away. But in the meantime, you can expect computers to take over more and more driving functions.

8 Grant July 18, 2010 at 3:01 am

Any autonomous vehicle relying on GPS isn’t going to be very useful. Tunnels, bad weather, tree cover, etc. can block GPS signals, and we can’t have cars flying off the road when that happens. At best they’d pull over and let a human take control in such circumstances.

…and then you still have to react to visible objects in the roadway that don’t have any sort of locator beacon. Cows, legacy vehicles, pedestrians, etc. All sorts of obstacles require visual processing in order to navigate.

What might be interesting is vehicles which drive themselves under certain circumstances, such as toll roads with a safe environment for them. Build fences to keep the deer and dogs out and only allow autonomous cars on the road. Then you could make sure that every lane, object and car on it broadcasts its location, and cars could locate themselves by a more local, land-based positioning system which wouldn’t be affected by weather. But is it worth the investment?

9 Andrey July 18, 2010 at 10:53 am

Google for “GPS jamming”, then think again about how awesome “hands off” GPS navigation on a vehicle would be.

10 Miller July 18, 2010 at 1:32 pm

Slocum,

You still haven’t explained why autonomous vehicles will require “turing-test level AI capabilities.” You keep going on about GPS. But GPS is not required for vehicle control. The primary purpose of GPS is navigation. GPS navigation is already in use in millions of vehicles. The main sensing technologies for vehicle control are radar, lidar and video. As I said, the DARPA Urban Challenge demonstrated that we can already build vehicles that can drive autonomously in urban environments, making complex maneuvers, avoiding obstacles, obeying traffic laws, and sharing the road with other vehicles. And that was three years ago. The technology has advanced even more since then.

A variety of driving automation technologies are already present in commercially available vehicles: ABS brakes, computerized traction control, drive-by-wire, parallel parking assist (computers identify a parking space and make all the necessary turns), lane departure warning and correction systems, active cruise control (which maintains a constant distance from the vehicle in front), pedestrian detection and avoidance systems, and a number of technologies for avoiding or mitigating collisions, like the Mercedes Pre-Safe Braking System and the Lexus Pre-Collision System.

Volvo expects to have an injury-proof car by 2020. GM expects to have a semi-autonomous vehicle by 2015 (the car will mostly drive itself in low-speed, congested conditions) and a fully-autonomous one by 2020. The primary obstacles to autonomous vehicles are social and political, not technological.

11 Slocum July 19, 2010 at 4:14 pm

You still haven’t explained why autonomous vehicles will require “turing-test level AI capabilities.”

Because vehicles are driven through environments that contain the complexity of the world. How often, for example, when you’re driving, do you try to determine whether or not ‘that other guy sees me’? How often to you recognize, for example, a type of car and a pattern of driving (aggressive young jerk in an SUV, old person who really shouldn’t be behind the wheel any more, distracted fool on a cell phone wandering around in and out of the lane) and predict their future behaviors and react accordingly? All the time, and this is something that is completely beyond AI systems. Now get off the highway and start driving through a busy city environment with pedestrians, cyclists, pets, kids playing, buses, garbage trucks, etc. Knowing what these things are and what they’re doing is important — and completely beyond current AI capabilities.

As I said, the DARPA Urban Challenge demonstrated that we can already build vehicles that can drive autonomously in urban environments, making complex maneuvers, avoiding obstacles, obeying traffic laws, and sharing the road with other vehicles. And that was three years ago.

This was still a very safe, controlled environment. It had other autonomous vehicles in the competition to avoid, but it didn’t have any of the other complexities of real urban environments that I mentioned. Get back to me when autonomous vehicles have been sent on a mission through a real city.

12 Miller July 19, 2010 at 7:12 pm

Slocum,

Sorry, nothing you write indicates a need for anything remotely close to “turing-test level AI capabilities.” Car safety systems already on the market can recognize dangerous patterns of behavior by pedestrians and other vehicles and respond accordingly. Automated systems can also measure speed and distance much more accurately than any human driver, have much faster reaction times than any human driver, and, unlike human drivers, never get tired or distracted.

And you’re wrong about the DARPA Urban Challenge too. The vehicles had to interact with one another and with human-operated cars driving around the same course and behaving as they would in a normal urban environment. The environment was not “safe and controlled.” About the half the cars were disqualified because they caused a collision or disobeyed a traffic signal or performed an illegal maneuver.

13 Slocum July 19, 2010 at 8:50 pm

I don’t see the problem here, either. There is nothing terribly challenging about creating a video processing system that can recognise a human police officer standing in an intersection and directing traffic.

Actually, there IS something terribly challenging about creating an artificial vision system that can reliably recognize a police officer standing in an intersection and understand the gestures he’s making–let alone comprehend and follow his spoken commands. AI systems aren’t close to that — digital camera face detectors work well enough to be useful, but only just–there’s an enormous gulf between those and what an autonomous vehicle would need.

And if all else fails, the human passenger could temporarily take control of the vehicle. In the scenario you describe, the passenger could simply direct the vehicle to pull over, or make a u-turn, or whatever else the cop told him to do.

Who’s to say there IS a human passenger — perhaps it’s an autonomous commercial truck hauling goods. And if there are passengers, who’s to say any of them are licensed drivers — maybe it’s a robotic school bus.

14 Miller July 19, 2010 at 9:52 pm

You’re still inventing non-existent problems. It wouldn’t need to “reliably” understand the cop’s gestures or spoken commands. As I said, if the automated system did not know how to respond, the human passenger could take control and direct the vehicle’s movement based on the cop’s instructions. If the vehicle was unoccupied, the cop could do the same thing. All it would need is some type of temporary manual intervention and simple interface to the vehicle’s control system. It could be as simple as speaking a short phrase (“turn left”) or tracing a maneuver with your finger on a touch screen, like an iPad. In fact, there are already GPS navigation systems on the market that are capable of understanding much more complex instructions than would be needed in this scenario. My own inexpensive Garmin unit allows me specify not only where I want to go, but the route I want to take and any detours I want to make along the way.

15 mbt shoes August 5, 2010 at 3:06 am

Who’s to say there IS a human passenger — perhaps it’s an autonomous commercial truck hauling goods. And if there are passengers, who’s to say any of them are licensed drivers — maybe it’s a robotic school bus.

Comments on this entry are closed.

Previous post:

Next post: