I agree that computers will soon (< 3 years) be able to beat most human drivers and as someone else said, getting rid of 70kg of driver is a huge advantage. For closed circuit courses like those used by Formula 1, the teams already have almost perfect topographic maps and models of the circuit, including measurements of the grip characteristics of the surface at various points on the circuit. Formula 1 teams run endless simulations and already know what the best possible lap should be.
But that's just qualifying. Racing is about so much more. Driving software will have to integrate considerably more inputs in order to avoid crashing into other cars. Let's say cars have to share telemetry. Development could probably happen in parallel with other improvements but I'd still stay 3-5 years.
Even if we assume software that can drive in a pack of cars, and algorithms for attacking and defending, there's that drive to win that makes some drivers do things that, if we're honest, aren't fair, safe or even right. The rules are different when nobody's life is at stake.
I would love to see Formula 0. No worries about driver safety. No artificial design limits to reduce cornering speed. In fact, the only design rule I would impose is the maximum height above ground and perhaps something about no laser cannons.
Does anyone remember Rogue and then Rog-O-Matic? It would be fun to create an online competition using an existing racing simulator.
> But that's just qualifying. Racing is about so much more. Driving software will have to integrate considerably more inputs in order to avoid crashing into other cars.
And that's where you see the difference between Gran Turismo's infamous zombie AI and Forza Motorsport drivatar (where everyone loves to hate M. Rossi). I bet that if the Turn 10 guys got in touch with those Stanford folks, they could learn each other a trick or two.
The BMW Track Trainer (http://en.wikipedia.org/wiki/BMW_Track_Trainer) can already beat most of the drivers. I think the difficult part is to beat the best ones. I doubt it will happen in the next 3 years, but I like the idea of Formula 0 very much :)
The thing I like most about the promise of self-driving vehicles is that, when they actually get to a point of real maturity, the cost to manufacture and put that tech into a production vehicle won't be significantly greater. We're talking about plugging into the ECUs, which already take metrics for vehicle dynamics to aid with ESC and other systems, coupled with what is, in essence, computers and IR sensors. The rest is just software.
In actuality, the self driving vehicle should cost us no more than $5k-$7k more on top of the cost of the vehicle.
There is unfortunately still a long way to go in cost reduction. The 360 degree LIDAR (Velodyne) that Google uses (the rotating thing on the top of the car) is the most essential sensor, but costs about $80000. Other autonomous cars also use quite expensive technology like laser scanners and so on. Unfortunately it is not as simple as only IR sensors :)
Not just driver error, but it reduces the need for everyone to own a car. Self-driving cars can move themselves to where they're needed, rather than sit in a garage waiting for their owners to use them.
The record is 10 minutes with purpose-built cars with titanic power (in the range of ~1000HP) and huge turbos to compensate for lack of oxygen at altitude. An Audi TTS with its measly 270HP can't beat that:
> According to Audi, race officials said an expert driver in car like the TTS would complete the course in around 17 minutes.
Have you ever watched humans race cars? I would watch robots race simple courses before I'd watch people, especially if different algorithms are competing. It'd be more like watching StarCraft bots compete.
Quite a lot of the population enjoys watching people race cars. Some people like football / soccer and I cannot fathom why, but I respect their choice of entertainment. I guess everything loved has a subtly that interests fans and bores everyone else.
I've had many a great race there back in the day. Seeing the lap times they knocked down is quite impressive too.
Thunder Hill is probably one of the more technically challenging tracks around. There are several off-camber turns, a nasty rise where you basically have to pitch yourself over it without being able to see (pucker factor 9,000) and two straightaways with flat and very oddly-aligned turns.
The fact that this TT was able to go so quickly around it is quite awesome - this gives me hope that driverless cars won't be as boring as I fear they'll be.
"It’s only a matter of time before robotic cars outperform their human-driven counterparts."
How much time are we talking? I don't agree with this statement in anything like the near (decades) term, particularly if we are talking actual races and not timed run events.
Actually, in some aspects, we're already there! :)
Computers can run a car at the limits of its traction circle much better than humans can, even with nothing more than an accelerometer. I've done some hacking around this, and it turns out that you can go about it fairly naïvely, code-wise, and do way better than even good amateur racers. Unfortunately, I don't have access to any top F1 drivers to datalog them side by side with my software, but maybe someday.
In terms of the quickest path around a given course, computers are _way_ better than humans at that. There's the official "racing line", which is a generally accepted fastest line, but as you can create more acceleration, the line actually flattens out towards the shortest route through a given section. Take a look at something like Spec Miata vs an AWD time attack car for a great demonstration of this.
Humans still have the edge on understanding & interpreting the environment. There's a lot of work being done in this area, though - I suspect we'll get some fairly robust stuff that's available to us on the open market in a few years.
Then there's the fatigue factor.
Heck, if nothing else, having to cart around 150-180lbs of meat with fairly tall packaging and some weird requirements about having to see and what kind of forces you can exert on it is a bit of a drag to making the quickest possible wheeled vehicle.
Sorry for the wall of text - this is something I've been hacking around a bunch for the last while, as it's the intersection of a bunch of my interests. Drop me a line if you want to chat more about it sometime.
"Heck, if nothing else, having to cart around 150-180lbs of meat with fairly tall packaging and some weird requirements about having to see and what kind of forces you can exert on it is a bit of a drag to making the quickest possible wheeled vehicle."
I think I should have gone into more detail, as I believe that the statement from the article refers to entering a computer controlled car into a race with humans.
The interaction is serious important, and knowing the fastest line doesn't mean your car will get to take the fastest line if someone is already there. Pitting is going to be interesting.
Some race series, NASCAR, only allow for a limited group of sensors. I would guess something like F-1 or Indy would be easier since they have more allowed telemetry. I would imagine a NASCAR driver robot is going to have a tough time telling the pit what exactly is wrong.
That in racing games the ideal line is the same for all car, much less for all suspension tuning, speaks to me of how utterly simplistic our models are for the best way to throw a car around the track.
The ability to reliably max out your traction circle is good, very last microsecond braking wonderful, and computers are wonderful at this micro-task. But knowing when exceeding it is ok, having a feel for where you need to be, and more than anything the planning and knowledge to max out turn exit speed velocity at all costs, these are all awfully complicated & sophisticated problems that have less to do with micro-optimizing the immediate present.
I certainly think cars will get there too. But I'd be surprised if we were really modelling the right problems at this stage. Emulating a person's sense of agency and purposefulness is, my guess would be, still a stretch. A person's ability to plan better is, my guess would be, trump enough over the micro-cheats of the machine.
It'll be interesting to see the domains where AI surpasses humans quicker. You mention the "no longer has to be a car advantage," which certainly holds true in the military drone regime, but where will in evened parity computerized elegance shine through brightest first? I tend to think something like F1, which is about extremely sharp reactions with extremely capable vehicles will lend itself to the micro-optimizations of the virtual well. What about a GT Touring race, or a club car race? Wouldn't it be a shocker if AI ends up crushing humans blithely in a stock Miata, but throw it in a GT car and it remains only human in performance? Maybe AI responsiveness and terrain estimation ends up dominating the rally circuit before the track? Mapping the domains where AI does succeed the most handily ought give the most remarkable contrast.
The fatigue factor really makes a big difference. In go, there are 5d (99th percentile) bots now, and towards the end of an hour-long game you can really see the human flagging.
Really? I'd expect that robotic cars will dominate closed-course races sooner rather than later simply due to the fact that the course is known, can be mapped down to the millimeter, and can be simulated offline. AI cars don't get bored, distracted, suffer fatigue, or need the physical stamina to drive an optimal line rain or shine hundreds of times in a row.
Mapping down to the millimeter is great, but the conditions on track are continually changing. Weather (rain, sun, track surface temperature, air temperature), debris (dirt, grass, marbles, spectator thrown shoes), rubber being laid down on the track, changing condition of the driver's vehicle and other vehicles are all things that need to be contended with. Simulation helps, assuredly, and is used in human based efforts extensively, but you can't exactly just play back your simulation on a real track and expect the car to stay on course at race speed.
That's a good point, for instance, Google's car has both on-board sensors and annotated route information. I wasn't suggesting that offline simulations be replayed, but simply that the problem space is to some extent a more perfect knowledge environment than the open road will ever be. Cheers.
Probably easier than open road, but the surface is a constant problem. NASCAR for example runs multiple races at some venues where the spring and summer races act totally different. Some of the differences are heat, rain between practice and race (rubber on track difference), different tire formula, possible resurfacing, and changes to rules (aero or engine).
Restrictor plate racing with a computer car mixed in would be extremely interesting. NASCAR does not allow telemetry, speedometers, or fuel gauges. This would make for an interesting computer car in these parameters.
Not a racing fan at all, but from friends I know there are certain strategies involved if each team consist of two or more members. One member can stall other racers, helping his team-mate advance. Again, this is just from talking to them.
Yep, happens in open wheel and drafting with teammates in stock car (although that isn't even totally true). It seems like it is much more prevalent where one sponsor sponsors multiple cars.
Thunder Hill has two nice long straights - you can generally top 120-140 with a decently powerful car.
What you need to knock down a quick lap there is brakes most of all. I had some monster Brembo GT's that allowed my little WRX to haul down from the triple digits and dive into the turns (paired with Tein coilovers). If you go up there you'll see how you have to cut in really fast towards the apex and then do a fairly ragged kick out the end to gain speed into the next straight.
Yeah, looking forward to Thunderhill in my BRZ. Infineon is fine in a supercapable car like an R8 but something with less grip is probably scary. (I got all wheels off in the R8 anyway)
But that's just qualifying. Racing is about so much more. Driving software will have to integrate considerably more inputs in order to avoid crashing into other cars. Let's say cars have to share telemetry. Development could probably happen in parallel with other improvements but I'd still stay 3-5 years.
Even if we assume software that can drive in a pack of cars, and algorithms for attacking and defending, there's that drive to win that makes some drivers do things that, if we're honest, aren't fair, safe or even right. The rules are different when nobody's life is at stake.
I would love to see Formula 0. No worries about driver safety. No artificial design limits to reduce cornering speed. In fact, the only design rule I would impose is the maximum height above ground and perhaps something about no laser cannons.
Does anyone remember Rogue and then Rog-O-Matic? It would be fun to create an online competition using an existing racing simulator.