Slashdot Mirror
Grand Challenge 1, Competitors 0
Ivan writes "According to the DARPA Grand Challenge Status Board, 2 bots were withdrawn before the race started and the remaining 13 were all disabled. Red Team and SciAutonics II tied at 7 miles, a bit short of the 142 miles required." CNN has coverage and interviews.
This was not a failure just because no one got further than seven miles. Contrary to a failure, this has been a grand success. DARPA spent around 13 million to host it, and got a lot of great minds in the public at large to start thinking of ways to solve very complex technological difficulties. In terms of sheer dollar value, the amount of technological research by private individuals easily surpassed the 13 million the government invested. Already companies are being created, and the wheels of commerce spun.
This benefits the public from the technology that is being created that otherwise lacked an impetus. It benefits industry by showing a host of new ideas that otherwise would have never come through the regular channels. It certainly benefits DARPA for sheer investment and public relations value. It can benefit future soldiers by reducing their risk to dangerous jobs. This also benefits the defense contractors that just got a small reminder that someone from out of nowhere could become a player - think of it as lighting a fire in their belly.
All told this was a challenge, and was never intended to be easily winnable. It certainly was advertised as being unlikely to be won this year. All told I think DARPA should hold more contests like this for other areas that have grown stagnant. For a historical perspective consider that Lindbergh crossed that Atlantic on just such a contest. A contest that inspired the X-Prize. Perhaps we should see DARPA become involved in future X-Prizes as well?
Just remember not to name the project skynet.
I don't think you've considered the immense complexity of simply adjusting your speed/direction to avoid a rock or pothole. Turn too fast, you flip (as at least one vehicle did). Next time you get in your car pay attention to just how many tiny speed/direction adjustments you make even on straight paved roads. Now add stuff you have to avoid and the process is incredibly complicated!
What's up with all the mechanical failures? Yeah, it's rough terrain, but we've been building human-powered vehicles that can handle it for decades! I'd think that keeping your engine going or your brakes from locking up would be the least "grand" part of the challenge.
You might to read the facts about challenge; SciAm for example had nice article.
But basically, it's not "just 142 miles in the middle of nowhere", but 142 miles with rather tight time limit (ie. they have to race almost as fast as human drivers would drive normally); exact route they HAD to follow (with some max. deviation allowed) was only disclosed few hours before start, and definitely wasn't just a straight line, and terrain was not just barren, it's pretty rough (meaning that staying on the road or path or whaver is a must) no matter how you look at it.
I like paying taxes. With them I buy civilization -- Oliver Wendell Holmes
If they were disabled due to "silly rule violating" then they deserve to be disabled. How would you feel if you were on one of the teams that got 7 miles and then had someone pass you and get 8 miles but broke 3 rules on the way.
They all knew the rules going in, and if they didn't comply then they deserve to be shutdown (this is ignoring DARPA's rule change to include teams that didn't finsih the qualifying course).
Normal people worry me!
I don't even know how to respond to that. I think your trolling but I'll respond anyway. Let's just say that driving is a much harder problem than you think and the course isn't all flat and straight. It requires vision systems to be able to compute the correct speed to take a cornor at, when to break, when to accellerate through a turn, when to compensate for loss of traction, stuck wheels, mud, sand, and mechanical failures etc. Oh, and the 1 m granularity of the mapping information isn't nearly enough to help compute the generally centimeter sized corrections that are required of a driver. Over simplifying problems you don't understand is a slashdot tradition I guess.
I wouldn't care if they got further breaking the rules, as long as they didn't win the prize if they finished the race. I look at it more of a showcase of technology than a showcase of rule-following.
I'm sure it'd be much easier to take a successful rule-breaking vehicle and tweak it to follow the rules than take a vehicle that couldn't get past 7 miles and make it a winner...
This is impossible to do. There are too many variables in the real world.
The bane of autonomous robotics is the fact you can't create an accurate world model. Sure, you can model the things you think will have the most effect, but there are literally millions of little things which by themselves may not mean much, but over time or in differing combinations can cause havoc and system disruption.
As an example, say for driving over barbed wire. Suppose you hit it at just such an angle the wire gets wrapped around the axle? You can't predict such things.
The post-mortems will be interesting to read. I hope they post them online.
-
I'll say this again. 1 m granularity isn't sufficient to make driving decisions with. Think of it this way. Could you drive in a city you've never been to w/o directions? No, well neither could a robot. Also, if you had a GPC accurate map would you drive w/o looking at the road? No, well neither could a robot. Oversimplifying problems you don't understand is a slashdot tradition I guess. Please tell me your not a London web site designer speaking about the state of AI research (like in the last ./ GC discussion)
Given that none of the teams finished, and in the proposed usage would allow such planning, I don't think this is a concern. Unless you don't like CMU for some reason, then you will complain until everyone stops listening to you.
There is an algorithm out there called the Kalman filter which does this. It's very complicated and rooted in probability theory, but it basically takes several sensor inputs, smooths out their response based on previous values (and known noise characteristics, such as the typical standard deviation from the truth) and makes a good assumption about where the sensors will be in the near-future.
It is very accurate, if you tune it properly (thats the tricky part)
This is very important for real time things because you need to begin to smoothly react to situations before they happen (ie, driving into an obstacle at high speeds).
-
Osama on a mule and foot is still eluding the US, the Iraqi guerrelas are either on foot or on mule, they continue to pester the US, whereas the Iraqi army, with trucks and tanks and other machines, was pulverized quickly.
... think of the Korean peninsula. If the north were to try to conquer the south, they would destroy its usefulness. Heh heh, if you want to think about something bizarre, think of the south surrendering as is to the north ... they would assimilate the north so fast, Dear Leader's head would spin as fast as his father in his grave.
Furthermore, empires today are built on economics, not military. It's bogus to even think of conquering western Europe, Japan, many of the small Asian countries, the US, Russia
Military might is only useful against dirt poor countries, and even then only in a limited sense.
Infuriate left and right
Whoa, look at this: with three minutes of typing and thinking, this Slashdotter just solved the problem that experienced engineers and computer scientists worked on nonstop for a year, at the cost of millions of dollars!
...
I have worked on a university project that is developing an autonomous robotic vehicle. The project is the Autonomous Robotic Vehicle Project (ARVP). Our team consists of a number of undergraduate and graduate engineering and computer science students. We participate annually in the IGVC (Intelligent Ground Vehicle Competition). The IGVC is a relatively simple challenge: navigate an obstacles course delineated by a pair of white lines marking the boundaries of the "road" and avoid obstacles such as traffic barrels. Max. speed is 5mph. Sounds simple right? Even a two-year-old could do it crawling.
Wrong. The IGVC has run for over 10 years, and in the last three years not a single team has completed the autonomous challenge. Teams from all over the world from prestigious universities compete, but autonomous navigation is not easy.
The first challenge is hardware: high-quality sensory equipment is expensive. Witness the thousands of dollars costs to purchase laser rangefinders, let alone radar or LIDAR. As well, computing power plays a large role, particularly at the speeds the Grand Challenge vehicles running.
The second challenge is software: It is not easy to write consistent vision algorithms. For example, simply the shadows cast off of trees on a sunny day can easily throw off line-recognition algorithms. I cannot imagine the difficulties that teams in the Grand Challenge experienced trying to recognize potholes and ditches.
The third challenge is reliability: All systems must work nearly perfectly together under a variety of conditions. The conditions at the Grand Challenge are far from ideal, and are in fact extremely difficult. The vehicles had to negotiate with dust, ditches and holes, overpasses, sand hazards, water hazards, fences and other obstacles.
All of these factors, which are quite easily dealt with in the amazing system known as the human mind, are very difficult for an autonomous vehicle. To be honest, I feel that seven miles is an extraordinary accomplishment under the race conditions. Few individuals honestly expected that the race would be completed.
Looks like most of the vehicles "crashed" (one way or another) pretty early on. Aside from a few scattered details (one apparently got tangled in barbed wire, a few flipped, some didn't start), anyone have a full list of What Happened to each of them?
TANSTAAFI: There Ain't No Such Thing As A Free iPod.
This is *great* news!
It means that autonomous fighting machines are still some way off.
This is a really weird sentiment to see on a technology website. I grant you that an autonomous fighting machine would be a bad thing to release on the world, but they'd still be a ways off even if some contestants passed the DARPA challenge. So many advances are necessary for an "autonomous fighting machine", that I think we can comfortably benefit from the development of robotic ATVs without worrying that they will someday rule the world.
Are you happy every time a chip design fails, because that postpones the inevitable rise of the "automous fighting machine"? Are you excited when you hear that Honda has to delay the release of ASIMO-2 because they can't get the hip-joints to work properly? Yet another set-back for the conquering strategy of the "autonomous fighting machine"!
It's also weird that someone else here thinks you're luddite comments are insightful.
Have teams done some real life testing before going to competition? Or they just did theoretical tests simulating on computer?
This is why new drivers and people who are disorientated or distracted often have trouble driving.
For new drivers there are 2 factors working here. First is the lack of experience of WHERE to look. New drivers often keep their vision trained too close in front of the vehicle. This works for very very low speeds but once you try to go road speeds you just don't have enough time to react. Experience teaches us to lift our eyes higher and seek ahead further in order to drive effectively.
The second factor for new drivers is how to handle this new kind of input. Stuff that you don't worry about when jogging or running becomes a big problem when driving, like corners and wet roads. The increased distance also means that you have to have a different sort of thought process in order to handle the increased amount of information.
With disorientated or distracted drivers they may have the knowledge of how to handle the processing of driving stimuli but since they are at diminished capacity they are not able to do so fast enough. Drunk drivers, for example, often start slewing back and forth because their reactions are lagging behind what their senses are telling them. They turn, overturn, correct, overcorrect, and so on.
Sapere aude!
The only obvious event that would jack his rating back up is capturing Osama. They have seven months to do it.
Note the increased military presence in Afganistan.
Current thinking (by us blessed madmen) is He Who Must Not Be Named has already be captured and will be paraded around the press 3 days before the election. Any longer and people will start asking awkward questions.
If the "awkward" questions they're currently asking HWMNBN get too tough, we'll likley see the body/remains "found" in a cave about 2 weeks before the election.
Yeah, I'm mad as a box full of spanners - and this post will disappear soon. But remember - you heard it here first!
Which brings us to the "so obvious, everybody's missing it" conclusion to the business of autonomous navigation over ROUGH terrain.
When these things are actually out in the field doing their jobs, they're going to be doing it on LEGS, not wheels, and they're gonna be going like a bat out of hell on top of it. Tell everybody you read about it first on slashdot if you'd like.
Should be interesting when the first legged competitor shows up for a Grand Challenge. Expect them to fail hilariously the first time or three, but after that, the laughter will turn into dropped jaws with no sound coming out.
Is it fascism yet?
It is true that passing a simulated test is no measure of success in the real world. But it will certainly be more prepared, and in a faster time and with less expense than an "all up" design method.
Look at the space shuttle if you want an example of "all up" gone bad. I'm not talking about the end product, I'm talking about billions that were squandered during development. The waste of time and money during the engine testing was extraordinary.
Another example is the Mark XIV torpedo. Google around, but the long and the short of it is the navy deployed a torpedo without testing it. A series of design flaws kept them from working, and their failure cost us dearly during the early parts of the war.
The Navy refused to believe there was a problem. The weapon worked 50 percent of the time for the 2 shots that were fired before the war. When they tested the torpedos properly they found numerous problems with the design of the guidance system and the detonators.
"Learning is not compulsory... neither is survival."
--Dr.W.Edwards Deming
Using a black-box approach, you could simulate the output of a "perfect" laser rangefinder, LIDAR, etc. In fact, black-box approaches are great for isolating bugs and system testing.
The point I'm trying to make is: if you limit the participants only to the well-heeled, you are not going to fully "unleash the entrepreneurial spirit", as was the stated goal of the competition.
A quote from a recent SBS Insight special on terrorism:
Have a good think about it. Why are they terrorists to start with? Maybe because they're sick of being fucked over by the US? People are generally peaceful and friendly until you put them under intense pressure. Ask a Palestinian about it.
If you hunt down 1 terrorist, another will arrive to fill in the gaps. If you are really serious about stopping terrorism, you need to attack the cause: the massive injustices in the world. Then terrorists will have no cause. Until then, expect more of the same.
After a few years competing in robot wars and battlebots I know how tough it is to think of something, pull it together in your spare time, and get it to the competition in one piece. And that's just souped up remote contol cars with saw blades. These are auto-fucking-nomous trucks.
This is damn good and all the competitors should be proud.
Providing everybody with a humvee is the wrong way to attack the problem. You are making the assumption that the humvee is the best platform to complete the challenge. I think the SciAutonics II team just proved that the humvee platform is not necessarily the best. They came up with a different platform that made also made it 7 miles.Perhaps the best platform has yet to be thought up. By specifying too much up front you could actually eliminate the best possible solution.
...how insane does this make the team that entered a motorcycle?
I suppose you weren't watching the live satellite feed when the motorcycle was demonstrated via remote control. It couldn't enter the race, but they just wanted to show it off.
It fell to the ground in literally 1 second.
Why they tried to solve a stabilization problem instead of an autonomy problem is beyond me. As I've said before, they engineered their own failing. This is different than the Red Team, where the basic hard problem of obstacle detection killed them.
Robo-Blogs of the world: UNITE!
One of those rules will eventually be "don't run over friendly troops". I don't think it's such a bad thing to force vehicles to stay within the rules. Otherwise we'll see over-optimization in areas which won't make sense down the road, pardon the pun.
Want to improve your Karma? Instead of "Post Anonymously", try the "Post Humously" option.
When these things are actually out in the field doing their jobs, they're going to be doing it on LEGS, not wheels
Actually, I think they'll use WINGS.
How's that for obvious?
Yah, flying is harder than driving. But autonomous flying over rough terrain is WAY easier than autonomous driving over rough terrain. Even autonomous landing is probaly easier than autonomous "running up a hill and jumping over a fence".
Most planes have been landed by computers for many years now. Sure it's on a smooth runaway. But throw in VTOL (vertical take off and landing -- which we've been doing for decades), and autonomous landing becomes way easier than driving.
You heard it here first: In the future, robots will fly.
" I suspect that the first industrialised nation that develops autonomous fighting machines will take over the world"
I predict the opposite. Any industrialised nation sufficinatly advanced to create an autonomous fighting machine would have little to gain from taking over the world. With adequate robot labor, you would have no need to exploit the world. At that point, added territory is no longer a source of useful resources but only an administrative burden. Primitive peoples are difficult to civilize and govern. Sure, we might use our robot warriors to down a particulalry bloodthirsty dictator from time to time and seed a self-governing democracy, just as we have used human soldiers to do with Milosovic in ex-Yugoslavia and Hussein in Iraq. But the goal in both places is to install a democracy and get the hell out ASAP. Fighting wars with robots will not change the underlying economic calculus of occupuation. It won't make ruling over the conquered any less of a pain in the ass, or any more profitable a proposition than today.
The more technologically advanced we become, the more we substitute common substances for exoctic mineral resource imported from abroad. Why conquer Brazil for copper mines when you get zillion times the bandwidth of copper from silicon glass fiber which is make from sand ? Power lines ? Use a superconductor strands. Conquer Africa for daimonds ? Bah !We can grow them more pure, large and cheaper in a vacuum deposition chamber in a New Jersey shopping mall. Once we find an adequate subsitite for fossil fuels, or choose to rely more heavily on those which we already have such as fission, that will be one less thing which we need from the outside.
The danger of autonomous fighting machines is not that the nations which develop them would use them to take over the world. The danger is that those weapons would fall into the hands of hostile and primitie societes which do have that goal, the same theat we face today. The technologically advanced nations which invented chemical, biological, and nuclear weapons could use them to to enslave the world, but we don't do that. The expense of occupation is too high and the payoffs for us are too low. The real danger of such technology is that is falls into the hands of primitive societies in which a primal warmongering mindset dominates.
Ceci n'est pas une signature.
The difficult problems need to be presented outright at first, so you don't invest too much time in something that can solve simple issues, but fails utterly at more complex ones.
The ultimate goal of autonomous robotics is to develop a system that interacts with the real world at least as well as human, if not better.
If you start off with a simple challenge, you will get simple answers. For the next challenge, you ramp up the challenge some, and most will just modify the simple system. At some point though, you can't modify what is fundamentally flawed, and you have to throw it all away and start over.
Thats a huge waste of time and resources. If the teams recognizes the *tough* challenges from the outset, they're more likely to come up with a system that is flexible enough to handle them when the time and ability comes. Granted, you may spend more time developing that framework before you solve simple issues, but its worth it in the end.
Now the teams know what real-world issues they face. Their future systems will be much better equipped to handle them as they come along.
I suspect DARPA was well aware that this challenge could not be met. But the teams and technology are better off for it.
-
I know exactly what you mean.
The interesting thing is to try and break it down to subroutines for AI programming.
First, there is the HUGE problem that everything we do, motor-skill-wise is completely unconsious. So all these weight shiftings and speed shifts and balance issues, they are all unknown territory for us. We are put in the position of having to guess what our own minds are thinking.
Then there is the whole "threat/obstacle" recognition bit. Human beings start developing that stuff long before we can walk, so by the time we could be running over uneven terrain we've got tens of thousands of hours of experience in this sort of thing, and every bit of that has honed our ability to recognise obstacles, and develop action plans.
All these advantages we have, and still, running over uneven terrain gives us that exilirating/terrifying brain-about-to-explode feeling.
Can you imagine what it would be like for a computer that has no experience with walking or with threat recognition to invent and reinvent them over and over again. And don't talk to me about learning systems; they don't learn very fast at all.
At this stage, it would probably be a better idea to build a machine that knows its own limits.
ad logicam Claiming a proposition is false because it was presented as the conclusion of a fallacious argument.
Lets think about what you've said. Obviously some of these teams are using state of the air equipment and technology. Any time you push the limits of what you can do you come away with data. I don't see how giving groups an incentive to do something outside of the lab is a waste of money. I also don't see how a "flashy race" as you call it is a bad thing. It is great to attack these problems from multiple angles and a flashy race is just one way of doing it.
Lasers Controlled Games!
You've hit upon a very big issue.
People don't work well together the way they used to. The open source movement is not an exception. These people all work virtually and at their own schedule and desire. It's very difficult to find committed people who can see the "big picture" without having to finance their loyalty.
A good analogy can be found in the music industry. What makes a great band often has more to do with X number of guys being open-minded and ambitious AND able to work well together. They may make a lot of mistakes and suck early on, but if they hang in there, they will prevail (look at Bon Jovi - talent is obviously not a prerequisite - tolerance is).