Slashdot Mirror
Teach A Robot To Drive, Win A Million Bucks
An Anonymous Reader writes "DARPA has released the details of a 'Grand Challenge,' with a $1 million prize. The challenge is to build an autonomous vehicle which can 'navigate on its own over a 250-mile desert course in less than 10 hours.' from L.A. to Vegas, 'without external communication or human control.' The contest is to be conducted in March 2004, and is open to all comers. Can we get at least one entry to represent slashdot?" We've mentioned this contest a few times before: any intended entrants out there want to disclose your secret plans?
All you need is to hook up a camerea up do an artificial neural net and spend a couple hours teaching/progamming it to steer and throw in a cpu and run a rule based system with a well defined set of rules(for navigation,traffic laws, etc) and we're set. Piece of cake.
I think about this problem off and on and I think that this project is conceivable. The key is to break down the software into separate pieces, each utilizing different types of computing. Mixtures of AI and traditional programming can be used to make this a reality.
My idea of the basic input system idea is based on layers:
Layer one: the camera(s)
two: various neural nets, each designed to filter out specific things. Fo example, one net would id cars (and their relative speeds), another the middle of the road (lanes), another the whole road. Maybe even one to find speed limit signs!
and tree: traditional algorithms which intepret the data from the neural net and use it to compute cars location, other cars locations, where the road is, what speed to go, etc etc.
This would allow the surrounding envornemt to be broken down into very simple data structures that traditional Algorithms can handle. I think the key to this problem is to divide and conquer, using the best tools for each part.
It should ot be difficult to train a NN to identify the boundaries of a road in the desert. The info from this can be transformed into 2d space and voila, you got a simplified but accurate view of the cars surroundings. Now just add NNs to id cars and you can use that for collision avboidance. I can go on and on, but you get the idea.
Its very annoying being a researcher with moral objections. DARPA is offering $40mil towards asynchronous research. Our research group is one of the best in the area but after discussing the situation we decided not to to take the money. Most of the group (including me) don't really want to do military research. One of the projects most suited to the group would be making some processors for missiles and I definately wouldnt be happy about that.
There is the case that we could do a pure research project for them. No direct link with weapons but simply making tools to make asynchronous processors but we may be blocked from publishing research and still (more indirectly) killing people.
I never though that in computers you would have to think so hard about what money and projects to accept but this is one I would skip. Its obveous that DARPA want some long range seeking technology but they want good engineers (ones who wouldnt work for them directly) to do their work and warm them over in the pretence that its a fun game. I can't think of many engineers who wouldn't want to have a go at this challenge.
Mouse powered Chips, Open source Processors and Lego
The robot ideally needs to be able to navigate without fear of being stuck on a 2" pebble 100' from the starting line. Obviously you can't map 250 miles of terrain down to millimeter resolution, but you can design the robot to render such obstacles nonexistant.
One concept is a large inflated sphere with light tread patches on the outside. The power/electronics pack is suspended inside with cables running to various points on the sphere. By adjusting the lengths of the cables, the sphere can shift the center of gravity and roll forward. A 6 to 10 foot sphere would allow most small obstacles to be avoided, then the robot only needs to note current position and the general surrounding topography.
Major difficulties with this concept are high winds (unless they are blowing in the right direction!) and steep uphill gradients.
...
3.1 Team Must Be U.S. Entity
The Challenge is open only to US entities. This includes U.S. corporations, U.S. non-profit organizations, U.S. universities, U.S. citizens, sole proprietors that are U.S. citizens or permanent residents, and partnerships of U.S. citizens or permanent residents.
It seems a lot of slashdot folk here don't read before they post. Some are already talking about avoiding other vehicles on the road. First off, do you really think they would allow a bunch of robotic vehicles to drive along side drivers? Second, it does mention across the desert both off and on road, but does not say anything about public roads where it must avoid other vehicles.
Another thing mentioned was GPS. Someone complained about not being able to use GPS. If that person had done some reading (FAQ on the page) then they would have seen that a public GPS or a private autonomous GPS receiver is allowed.
Now my thoughts on this challenge.
A few things I think are most important here are:
4x4 type of vehicle (truck, SUV, Hummer, Jeep, etc.) (Automatic prefered for ease of use)
A must have GPS receiver.
Infrared obstacle detection device (180 degrees) about 100 meters.
Also, another device to analyze the terrain about 180 degrees around the front of the vehicle out to about 100 meters.
Attach the GPS, obstacle device, and terrain device to a computer and also have the computer hooked up to the acceleration pedal, brake pedal, and shifter to put into park or drive.
The terrain device would be the most complicated. Then all you would need is a few good programmers that can work with the data the different devices provide and your set. I'm not sure if such terrain devices exist outside the military, but I'm sure some laser/infrared/etc. engineers out there could produce a basic one. This project isn't as complicated as it sounds, but it would take some good engineers and programmers to finish. Just my thoughts.
Question everything.
According to the article:
Examples of obstacles include ditches, open water, rocks, underpasses, and construction. All obstructions on the route can be either accommodated or avoided by a commercial 4X4 pick-up truck.
Anyone that has gone offroad can attest to how hard it is to go 25mph consistently or on average. Add in to that the fact that you will have to backtrack and figure out alternate routes means your vehicle is going to be thinking fast and driving very fast.
The article mentions part of the route will be on paved roads, so maybe you can make up a lot of time on those stretches.
Definitely not a trivial challenge. but a fun one for sure. Some of the non-trivial ingredients: the offroad vehicle, gps for detecting way points and finish points, camera for detecting obstacles and terrain, algorithm to determine fastest speed for current terrain, feedback to determine whether you are getting close to tip over, algorithm for determining alternate routes.
what else?
G'morning all.
:)
:) Radio stations make for decent markers too. Since you have something the size of a truck, it would be a piece of cake to triangulate distance and direction to any one becon, and use any two to fix location.
:)
This sounds like fun.. Personally, I don't have *ALL* the skills required to pull this one off, but if anyone's forming a team in the Los Angeles area, I'm in.. I have skills everywhere from the technical aspects of making a vehicle work to engineering of the hardware involved.
I'm thinking something like a slightly modified S-10 Blazer, or K5 Blazer. Positraction (not available on the older S-10's) is a must. Probably the K5 would be the better choice, for extra room in the engine compartment for controls.
I did a quick read through their forums. There's some interesting (and optimistic) talk of stereo vision through, laser vision/guidance, and ground evaluation through radar.. A few of the people sound like they have a clue, and some others didn't even read the rules..
Some of them are talking about exotic hardware solutions, that they'll probably spend all the available time building, and then wonder why they don't have a working vehicle to go with it. Some others were talking about cool Xeon based systems, and forget that they get hot, and this is going to be running in a vehicle in the desert for 10 hours. One mentioned the hardships of hard drives, and doesn't even realize that you can use Compact Flash as your hard drive, and do stuff from there. No one yet mentioned using Linux..
My thoughs on a practical vehicle is a late 80's Chevy K5 blazer. Radar (like the backup radar in late model Lincoln's) to evaluate for local blockages. Vision system, like a stereo camera hooked up to a Linux box (this is where I'm at a loss. I don't think I could do this software).
Steering control would be an electric motor with chain drive just before the steering box. That way, no major changes to the steering need to be done.
Acceleration is a simple motor pulling on the throttle assembly, just like the vacuume accuator on cruise control.
Braking would need to be something more substantial. probably a pneumatic ram on the brake pedal lever itself.
I'd suspect it'll take a few computers to run it, but in something the size of a K5 blazer, we'd have no only room to mount it, but more than enough room to mount it preventing shocks... The computers would need to be hard-drive free though.. Compact flash cards of say 512Mb would be just about all we'd have to work with. That should be sufficent though.
The site says they're providing several checkpoints which are mandatory to pass through/stop at. There will also be mandatory waypoints, which define the path. Fairly easily, go from waypoint to waypoint. If there's an obstical, decide for left or right turn to go around.. More than likely the easiest thing would be to use GPS to establish a location (when available), and use other public navigation beacons the rest of the time.
Anyone who's flown knows how many radio navigation beacons are available.
When you detect an obstical, mark it on an onboard map, and figure out a way around. That would be for big obsticals like canyons or mountains. Small obsticals, you steer around.
I can design and build anything required to make the vehicle itself work. Navigation will be up to someone else. This is/will be a team project, so as many hands as we can get involved would be cool.
Can we get Cmdr Taco's permission to put "Slashdot" down the side of the truck?
Who's in? Reply here first, then we'll get in contact in real life.
Serious? Seriousness is well above my pay grade.
I have on good authority that the Jet Propulsion Laboratory has a huge lead over most competition. They already have autonomous vehicles that can travel at ~15 mph over very rough terrain. Keep your eyes out for the Caltech/JPL team to make some real waves.
I'd love to work on something like this, but I don't have the money so all I can do is "armchair quarterback" it. Assuming money was no object, and I had access to all the fun toys that are currently available....
- Vision: synthetic aperature radar. This would be the ideal way to detect potholes, judge distance and height, etc. Very expensive stuff, unfortunately. Also include forward-looking color camera to read signs etc.
- Processing: one poster suggested "divide and conquer", which would probably be the best approach. Feed the world data from the radar/camera into multiple systems, each for dealing with a different problem. Here's what you'd have to be able to deal with:
1. identifying road positions. Not too hard to do, just look for the center line and shoulder.
2. identify unpaved road paths. (gravel and dirt) More complicated.
3. Identify road hazards.
4. Process road signs. (speed limit in particular)
5. Navigation during maneuvers. (merging, turning, changing road types, stopping, pulling over and pulling out)
6. several smaller algorythms to detect uncommon but possible conditions, such as encountering a slow moving vehicle, cattle on the road, being passed by another vehicle, identification of blocked roads (construction, bridge out, detour) and so forth. Each of these would have to be handled as an interrupting action.
A GPS will no doubt be essential to any design.
A fully loaded map system would have to be in the mix as well. Think MapQuest or something like that. It would need the ability to plot a sensible course from any two given points.
Navigation would combine all of these elements. Use of the GPS and the road detectors would keep the "bot" on course. Road hazard system would interrupt to deal with dynamic situations and general vehicle safety. In the event of a road hazard, it would probably be smartest to simply mark that bit of the road as "no longer there" and calculate a new route. Better to turn around and detour to go around a tumbleweed in the road than to try to run it over.
Speed determination would require some thought. Road conditions, weather, terrain roughness, speed limit, would all have to be factored in.
Dealing with various traffic controls would be necessary. Stop signs, speed limit changes, etc would be not challenge too much.
Uncontrolled intersections would be a pain, and might require the radar/camera to be able to swivel around to "look both ways before crossing" etc.
Off-road conditions would be especially difficult. Obstacles would be difficult to quantify. (just how much do we need to slow down for that pothole?) This is why the radar would be so important... stereo vision is too easily confused by shadows and other factors, and if you're trucking along at 40mph you have to know if that's a shadow or a hole or a small boulder.
It would probably be a good idea to have some reasonable additions made for general vehicle monitoring. Gas, temp, oil pressure, interior temp, spedometer, tach all would need to provide feedback to the system. It would be a good idea to have a way to detect a flat tire, but I can't think offhand of a good way to do that, especially considering the possibility of needing to cross rough terrain.
Hopefully they would allow some reasonable safety controls. I'd like to see a live camera feed to monitor progress, and a set of like three kill buttons that can transmit emergency overrides in the event of malfunction. One to trigger an orderly "pull over and stop", one to do an emergency stop (hard brake and kill engine) and one to "kill engine and immediately shut down".
Lastly, an "overseer" would be a very wise thing to include. A small computer running on independent power with interrupt control over the steering, throttle, brake, and ignition. It's only job would be to monitor the navigation system for faults. If navigation goes out, the overseer would quickly but safely stop the vehicle, and attempt to restart the navigation. (reboot) If this fails, it would stop the ignition and trigger a distress beacon.
I work for the Department of Redundancy Department.
The Navlab vehicles do not qualify because of where the funding for the project came from. I can't remember the details but it basically disqualifies any Navlab vehicle (at least that's the last I heard). So, CMU will either not be attending, or else someone will put together a new vehicle, and we'll kick everyone's butt, just like we do at robotic soccer.
For German carmakers this is already state-of-the-art in terms of prototype cars. For example the is development at DaimlerCrysler with a self driving car. During my practical back in 1996 this car was already able to drive by its own (first test on an unused airport lane). Currently they are so far that this car is able to detect unexpected behaviour an to react accordingly, e.g. playing kids that runs on the street.
Another development is at Volkswagen they have already a self-driving car that runs on their test course. This used for automatic testing as far as I know. This car was quite new about two years ago.
But for any of these cars they have to run on a road, because besides the positioning they need the marking of the roads.
TS