How long will it be before some clever person programs the Tesla self driving feature to do autocross laps at the absolute limit of speed? We know humans are flawed but having the computer do the driving on the absolute edge should pretty much guarantee a win.
Never
If you're actually programming in the course, it could probably be done now. If you want the computer to find it's way through a cone maze, there's still a lot of work to be done.
I'd rather see driverless F1 racing though, get the squishy sensitive meat bag out from behind the wheel and see what kind of performance they can pull when you don't have to worry about the driver blacking out from G forces.
RevRico said:If you're actually programming in the course, it could probably be done now. If you want the computer to find it's way through a cone maze, there's still a lot of work to be done.
Agreed, reason I can't autocross is because I cant find the path around the course in the first place; forget thinking two corners ahead.
If you were to drive around the course slowly, and have the GPS program the path, I believe the car could navigate just fine. But you would need to map the nuances of the surface and track if the thing was to be competitive.
NOHOME said:Agreed, reason I can't autocross is because I cant find the path around the course in the first place; forget thinking two corners ahead.
And you've probably got a big advantage over the Tesla there. Here'a a case where Tesla's Fool Self Driving came to a fork in the road, couldn't decide which side to take, and split the difference.
In reply to NOHOME :
It probably would take a little AI to optimize final speed. If you look at the inputs in Formula 1 from data sent one way I have no doubt the teams are indicating an optimum corner speed and braking threshold for the driver to hit based on projected life of the tires.
MadScientistMatt said:NOHOME said:Agreed, reason I can't autocross is because I cant find the path around the course in the first place; forget thinking two corners ahead.
And you've probably got a big advantage over the Tesla there. Here'a a case where Tesla's Fool Self Driving came to a fork in the road, couldn't decide which side to take, and split the difference.
Rumor has it there aren't any forks in the road on an autocross course. Besides if they had them everyone would be getting flat tires.
In reply to frenchyd :
See, that is where you are wrong. All I see on a cone course is forks in the road.The gap between every single cone is an option that needs to be processed for the possibility that it may or may not be the correct path. I think AI would have the same issue to overcome.
Since the main point of autocross is the driver, self driving cars don't make much sense. Now remotely driven cars could be a possibility, but still not something I would prefer.
frenchyd said:Rumor has it there aren't any forks in the road on an autocross course. Besides if they had them everyone would be getting flat tires.
Every autocross with a crossover or multiple laps has forks in the road. Which one you take depends on where you are in the run or what lap you are on.
On the flip side - Teslas can see traffic cones really well, so the usual problems with recognizing weird intersections and things like cardboard boxes don't happen. And to be on a level playing field with the meat driver, you should let the car take a guided slow speed tour of the course before the competitive runs. So, we have a car that can easily see the cones and knows where the course goes thanks to a high resolution GPS track - it's a much easier problem to solve than navigating a city center.
IIRC when Loeb set the record at Pike's Peak, he was able to beat the calculated ideal time by a handful of seconds.
81cpcamaro said:Since the main point of autocross is the driver, self driving cars don't make much sense. Now remotely driven cars could be a possibility, but still not something I would prefer.
We sure spend a lot of money on our cars if it's all about the driver :)
Self-driving cars are built with "machine learning" technology, and one of the attributes of that tech is that the more narrowly defined and controlled the problem is, the better the tech works at solving it. They've already got computer-controlled cars that can lap race tracks with very respectable times, and I imagine it wouldn't be all that hard to adapt the software to autocross. Feed it the course map ahead of time and give it a lidar system and it can probably see its way through the "sea of cones" better than most people can because it can constantly refer back to the map. GPS helps too.
Feeding course walk info into it would be trickier, you'd probably need a human to program in information about surface imperfections, bumps, etc. There are also a few corner cases that might be trickier to train, like stopping for downed cones and how to not mow down corner workers who are in places they shouldn't be.
I suspect that we've already got the tech we'd need to build a self-autocrossing car, and with an appropriate budget it could be done. It'd probably be very fast, potentially faster than any human driver.
They already make a self-driving race car that emulates me on a track:
Surface imperfections would be easy enough to learn if you had suspension position sensors - not all that rare. Even the course walk wouldn't be that hard, just have someone walk through the course with the right instrumentation. I don't think it would be all that difficult as these things go. Making life easier for the car is the fact that, unlike a permanent race track, the meat drivers don't have time to completely perfect their lines through an autox course. They might get close, but that last x% isn't there by nature of the short-lived course. So the car doesn't need to get perfection either.
I think downed cones and random pedestrian detection would be outside the bounds of this exercise and would be moving towards full service autonomy instead of "self driving autox" - as noted, that gets much harder to solve very quickly. There aren't supposed to be downed cones on a properly run autox course. Same with moving targets. We would also have to program what happens if a helicopter lands in the middle of the course :) Although being able to compare cone positions with a high level of accuracy between runs would make the downed cone detection relatively simple. Cars have much better memories than people.
The really fun thing would be if you had two or more automonous cars running as a "team". Meat drivers talk to each other and watch each other. The fun thing about the cars is that they'd all learn what the others had learned so you could potentially have a multi-way tie for FTD :)
Keith Tanner said:Surface imperfections would be easy enough to learn if you had suspension position sensors - not all that rare. Even the course walk wouldn't be that hard, just have someone walk through the course with the right instrumentation. I don't think it would be all that difficult as these things go. Making life easier for the car is the fact that, unlike a permanent race track, the meat drivers don't have time to completely perfect their lines through an autox course. They might get close, but that last x% isn't there by nature of the short-lived course. So the car doesn't need to get perfection either.
I think downed cones and random pedestrian detection would be outside the bounds of this exercise and would be moving towards full service autonomy instead of "self driving autox" - as noted, that gets much harder to solve very quickly. There aren't supposed to be downed cones on a properly run autox course. Same with moving targets. We would also have to program what happens if a helicopter lands in the middle of the course :) Although being able to compare cone positions with a high level of accuracy between runs would make the downed cone detection relatively simple. Cars have much better memories than people.
The really fun thing would be if you had two or more automonous cars running as a "team". Meat drivers talk to each other and watch each other. The fun thing about the cars is that they'd all learn what the others had learned so you could potentially have a multi-way tie for FTD :)
By surface imperfections I meant the way that human drivers will use information gleaned from the course walk to modify the line they would otherwise take through the course ahead of time. Yes, the self-autoxing car can sense those bumps when it drives over them, but it's used up one of its runs to do that rather than knowing about it ahead of time.
Downed cones are definitely a part of the sport -- the intention is that corner workers get them all set up again before the next car comes around, but that doesn't always happen and being able to spot them, stop before them, and demand the rerun is important. :) As for "pedestrians", if you wanted to run a self-autoxing car at a normal autox event it would need to be able to safely negotiate a course where corner workers are standing in (presumably) safe spots, and be able to detect when those corner workers are out of position in a dangerous way and stop before it gets to them. Realistically, I suspect the liability associated with this is where the idea stops.
Perhaps instead of self-autoxing cars what we need is to build robotic self-resetting cones!
Here's a better question- why does it even matter?
Machines have been able to beat the best chess players in the world for many years now, but people still play chess.
Which means if a machine is able to be the fastest driver ever all the time, it really doesn't matter- people will still autocross. Because fun.
81cpcamaro said:Since the main point of autocross is the driver, self driving cars don't make much sense. Now remotely driven cars could be a possibility, but still not something I would prefer.
Oh, I'm sure a spacer would be in the seat and could do the human stuff outside the limits of the programs. ( cone, out of position meat, etc) the speed trick would be to get the lightest legal spacer.
alfadriver said:Here's a better question- why does it even matter?
Machines have been able to beat the best chess players in the world for many years now, but people still play chess.
Which means if a machine is able to be the fastest driver ever all the time, it really doesn't matter- people will still autocross. Because fun.
People still ride horses.
Toyman01 + Sized and said:frenchyd said:Rumor has it there aren't any forks in the road on an autocross course. Besides if they had them everyone would be getting flat tires.Every autocross with a crossover or multiple laps has forks in the road. Which one you take depends on where you are in the run or what lap you are on.
Shouldn't be any problem to program. I was trying to be funny
alfadriver said:Here's a better question- why does it even matter?
Machines have been able to beat the best chess players in the world for many years now, but people still play chess.
Which means if a machine is able to be the fastest driver ever all the time, it really doesn't matter- people will still autocross. Because fun.
But they only play chess against other humans, or against handicapped machines :)
It's a fun thought and programming exercise, and it expands our understanding of how to do this in other areas. But mostly it's a challenge.
frenchyd said:alfadriver said:Here's a better question- why does it even matter?
Machines have been able to beat the best chess players in the world for many years now, but people still play chess.
Which means if a machine is able to be the fastest driver ever all the time, it really doesn't matter- people will still autocross. Because fun.
People still ride horses.
And drive them, thank you very much.
In reply to Keith Tanner :
If it's more fun to someone to have an automated car do an autocross, fine with me.
I almost never get a chance to go cone dodging, but I can try to solve interesting programming problems all day long :)
You'll need to log in to post.
