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?

Comments

Hrmmmmm

[Anonymous Coward]

I was jkust talking with a fellow from the same university I attend, and there is a group that is has done this for a one mile range.

(It's a robot that finds a specific building within a one mile radius and does other things involving the building)

I hear they've got the "flying to a building within a 1 mile radius" part done. Wouldn't it be not much more difficult to extend the radius to 250 miles? What would be involved?

Re:Sounds cool...

[m_xiphias]

There's E-Stops built into all of them. A chase vehicle is required, with a judge onboard. If the chase vehicle is too far away or something bad is about to happen, the E-Stop will be activated and the robot must come to a safe stop. At least, that's how it works in theory. Obviously, DARPA is trying to be as safe and thorough as possible.

Perhaps not that hard?

[ryants]

This is from Artificial Intelligence: A Modern Approach, by Stuart Russel and Peter Norvig, published 1995:

ALVINN (Autonomous Land Vehicle In a Neural Network) ... is a neural network that has performed quite well in a domain where other approaches have failed. It learns to steer a vehicle along a single lane on a highway by observing the performance of a human driver. ... The results of the traning are impressive. ALVINN has driven at speeds up to 70 mph for distances up to 90 miles on public highways near Pittsburgh. It has also drive at normal speeds on single lane dirt roads, paved bike paths, and tow lane suburban streets.

The only problem is the training... the system is unable to drive on roads that it doesn't have training data for. I glanced quickly at the DARPA rules and didn't see anything that would invalidate a "build a similar course and train on it" approach. So take ALVINN, build lots of courses that sound like the sort that DARPA is planning, and train, train, train!

References:

• Artificial Intelligence: A Modern Approach
• ALVINN

Re:Perhaps not that hard?

[Anonymous Coward]

ALVINN is one of several subprojects of the long running NAVLAB project at CMU. I took classes from the professors working on this project and XAVIER(a robot that can navigate halls teaming with people w/o bumping into them). They use multiple systems based upon different approaches(neural nets, bayian, etc.) and the systems vote to decide the correct decision. It don't have to train upon a given path, but upon any road that is of similar terain(ie just train the system upon desert roads). I don't want to discurage anyone, but CMU has been working on this problem for at least 12 years and the seven or so generations of the system have been progressing at an impressive rate. The system can currently run on a laptop! They have the inside track and I think most of their current funding comes from DARPA already. Good luck to anyone working on this, you'll need it. Driving on rough terain is a tough problem, because GPS isn't accurate enough for driving(only general direction finding).

Re:Somewhat ridiculous requirement....

[larien]

The FAQ states that GPS is specifically allowed.

Even Faster Than That

[sacdelta]

From the rules:

2.19.3 Maximum Finishing Time

In order to qualify for the Grand Challenge cash award, the maximum corrected finishing time of the winning team must be less than six hours. Additionally, to ensure safe operation during daylight hours only, all vehicles must be removed from the route ten hours after their departure.

If you want the money you have to do it in 6 hours. This is just over 40 mph. But some of the route is paved which should allow for higher speeds for parts of the course.

The rules also state that the route will be navigable/avoidable by a standard 4x4 pickup (HINT HINT).

Having driven out in the desert, even on the dirt roads, most of it can be driven at 60+ mph. You just have to be ready for the parts that can't be driven faster than 5 mph.

Re:Even Faster Than That

[sacdelta]

2.24 Challenge Area

The Challenge area includes the Departure Area, Departure Line, Challenge Route, Checkpoint Area, Arrival Line, Arrival Area, and any other area that has been assigned to DARPA for the purpose of conducting this Challenge. The specific boundaries of the Challenge area will be briefed to the Participants at a pre-Challenge brief shortly prior to the Challenge.

It sounds like they give the details of the route only a short time before the actual race (maybe a couple of days? hours?) so it would probably be a good idea to have built-in GPS to assign the waypoints quickly, easily and accurately.

The point of the exercise is to see how well the robot car can deal with "unknown" conditions, so I would wager that pre-driving would be discouraged.

In fact:

Vehicles that cannot demonstrate intelligent _autonomous_ behavior will not be accepted as Participants.

So a marker based vehicle would be right out.

Re:Uh, riiight....

[effer]

From the FAQ...

Q11. Can I use differential global positioning system (GPS)?

A11. The challenge vehicle is free to use publicly available signals. This includes differential GPS receivers in towns or counties along the way. A team may establish a private differential GPS receiver, as long as it is fully autonomous, at a checkpoint.

Re:Team Slashdot?

[AsOldAsFortran]

There's lots of starter code for computer vision tasks for those who want to play with the project. A major web page with links to research groups and software sources is

http://www-2.cs.cmu.edu/~cil/v-source.html

If you want to use intel processors, then there's a library of routines for many vision tasks supported by Intel at:

http://www.intel.com/research/mrl/research/opencv/

Here's what's fun about opencv - it's an effort to make vision algorithms as accessible to programmers as opengl makes graphics algorithms.

Comments by an entrant

[Animats]

We're entering this. A few comments.

• You can't game your way around the rules. You have to describe your approach to DARPA in writing, and DARPA reserves the right to change the rules after entrants have submitted their technical specs. They want a useful autonomous vehicle, not a trick.
• The rules have changed several times, and will change again. There's supposed to be a more or less final version on 1 April 2003. Right now, the announced plan is Barstow to Las Vegas in 10 hours.
• You can't preplan the whole run using map data, aerial imagery and GPS. DARPA will do things to make that not work, like placing some obstacles on the route. Note that you get the route, in the form of about 1000 waypoints, two hours before the race.
• DARPA does not guarantee that the course will be cleared of other persons and vehicles. Early versions of the rules said that the course would be cleared, but then DARPA changed the rules. Now it's only a "best effort" thing. Some competitors pulled out at that point. There will be sweeps ahead of the robot vehicles, vehicles following behind with remote emergency stop buttons, and road closures, but somebody still might not get the word. The route isn't on military bases; it's on Bureau of Land Management land open to the public. DARPA claims they will come up with an insurance carrier that will provide liability coverage, but so far, that hasn't happened. Vehicles thus need very good safety systems.

It's a nice engineering challenge. All of us have solved tough problems in the past, and we've all done serious robotics work. This looks within reach, although difficult.