ROBORace

[SR71]

http://www.theverge.com/2016/8/22/12592 ... type-watch

[FW17]

Hope it has a kill switch

[Andres125sx]

From SR71 link:

It sounds like we could know more soon. Roborace's DevBot will make its public debut on August 24th, at the Formula E open practice sessions at Donington Park

Any new?

[Jolle]

Actually, I think driving a car on the edge on a track (just one step away from AI racing) isn't that hard of a challenge as many of you think. It's mostly pure set action-reaction, drivers react almost instinctly on input received like oversteer or a locking wheel. Many of those action-reaction parts of driving a car in anger are already in the hands of systems, like ABS and TC. Even with possibilities that go beyond our capabilities like ESP.
Next to braking and accelerating there are basically just corners, which ESP already knows lots about.

I think, if they put their mind to it, a robot would lap faster then Hamilton (especially over a GP distance).

[Pat Pending]

Actually, I think driving a car on the edge on a track (just one step away from AI racing) isn't that hard of a challenge as many of you think. It's mostly pure set action-reaction, drivers react almost instinctly on input received like oversteer or a locking wheel. Many of those action-reaction parts of driving a car in anger are already in the hands of systems, like ABS and TC. Even with possibilities that go beyond our capabilities like ESP.
Next to braking and accelerating there are basically just corners, which ESP already knows lots about.

I think, if they put their mind to it, a robot would lap faster then Hamilton (especially over a GP distance).

IMO, yes and no. I agree that controlling understeer/oversteer and dealing with wheelspin/locking wheel is easily done with current ESP, ABS etc. I think where things are more difficult is in the 'vision' side of things - detecting the edges of the track, 'seeing' if the upcoming corner is an easy curve or 'throw out the anchors'hairpin etc, plus picking the correct racing line & knowing when it is fact quicker to take a 'non-optimal' line for one corner in order to be quicker through a following corner, picking braking points etc.

I can see two different approaches to that just myself; i. throw computing power & sensors at it and try to work it all out in advance of arriving at a bend, or ii. iteratively adjust speed, brakes etc lap after lap to find best parameter values. Or a combination of both, more likely.

All that is impressive enough with 1 car on a track but that ain't racing. The really, REALLY, difficult bit will be dealing with all that stuff AND avoiding other cars on the track. But that STILL ain't racing, is it. Racing is consciously braking late in order pass another competitor, for example. Be interesting to see how that decision making is implemented.

[Jolle]

Actually, I think driving a car on the edge on a track (just one step away from AI racing) isn't that hard of a challenge as many of you think. It's mostly pure set action-reaction, drivers react almost instinctly on input received like oversteer or a locking wheel. Many of those action-reaction parts of driving a car in anger are already in the hands of systems, like ABS and TC. Even with possibilities that go beyond our capabilities like ESP.
Next to braking and accelerating there are basically just corners, which ESP already knows lots about.

I think, if they put their mind to it, a robot would lap faster then Hamilton (especially over a GP distance).

IMO, yes and no. I agree that controlling understeer/oversteer and dealing with wheelspin/locking wheel is easily done with current ESP, ABS etc. I think where things are more difficult is in the 'vision' side of things - detecting the edges of the track, 'seeing' if the upcoming corner is an easy curve or 'throw out the anchors'hairpin etc, plus picking the correct racing line & knowing when it is fact quicker to take a 'non-optimal' line for one corner in order to be quicker through a following corner, picking braking points etc.

I can see two different approaches to that just myself; i. throw computing power & sensors at it and try to work it all out in advance of arriving at a bend, or ii. iteratively adjust speed, brakes etc lap after lap to find best parameter values. Or a combination of both, more likely.

All that is impressive enough with 1 car on a track but that ain't racing. The really, REALLY, difficult bit will be dealing with all that stuff AND avoiding other cars on the track. But that STILL ain't racing, is it. Racing is consciously braking late in order pass another competitor, for example. Be interesting to see how that decision making is implemented.

Well, just like a human driver the bot has to lern the track. And track limits? With a bit of sensors or even beacons that's relatively easy. I think most sims of the teams can drive an optimal lap on their own quite easily, then it's just translating that to a drone.

On real wheel to wheel racing, that's the hard part in my opinion but not extremely difficult. The rules of engagement are good translatable into "if-then" code, could even incorporate some chess like algorithms to calculate the outcome of certain actions and making strategic decisions.

The beauty of a race track in this case is that it's a close situation. It's relative simple. Everybody is going the same way, no other trafic, very little change in variations etc etc.

[Andres125sx]

I think where things are more difficult is in the 'vision' side of things - detecting the edges of the track, 'seeing' if the upcoming corner is an easy curve or 'throw out the anchors'hairpin etc, plus picking the correct racing line & knowing when it is fact quicker to take a 'non-optimal' line for one corner in order to be quicker through a following corner, picking braking points etc.

I don´t think the ´vision´ part will be difficult, they can use tons of sensors for that, and they can be placed higher for better view, unlike driver´s eyes wich must be low in order to lower CoG. Also, proximity sensors are a lot more accurate than a driver evaluating if his front wing will hit the car in front or not.

To me the most difficult part will be programming unusual situations. For a qualy lap I think it must be quite easy to program a drone. Actually Audi already did it. The difficult part will be to evaluate when the ideal line is not the ideal line anymore because there´s a car, or because it´s wet and the rubber makes the ideal line slower. I think that decisions are what will make the difference

Also, I wonder how a computer will evaluate when braking a bit late is worth to overtake a car or not. This is standard procedure for human drivers, if you go a bit long but not enough to go out of track so you can keep track position, it´s worth even if you did the corner slower than posible. But if too long, even if keeping on track, the car you just overtook will pass you again. Where´s the limit? Programming those limits will be a key factor.

BTW, to stablish those limits I hope they don´t ask Rosberg

[Jolle]

I think where things are more difficult is in the 'vision' side of things - detecting the edges of the track, 'seeing' if the upcoming corner is an easy curve or 'throw out the anchors'hairpin etc, plus picking the correct racing line & knowing when it is fact quicker to take a 'non-optimal' line for one corner in order to be quicker through a following corner, picking braking points etc.

I don´t think the ´vision´ part will be difficult, they can use tons of sensors for that, and they can be placed higher for better view, unlike driver´s eyes wich must be low in order to lower CoG. Also, proximity sensors are a lot more accurate than a driver evaluating if his front wing will hit the car in front or not.

To me the most difficult part will be programming unusual situations. For a qualy lap I think it must be quite easy to program a drone. Actually Audi already did it. The difficult part will be to evaluate when the ideal line is not the ideal line anymore because there´s a car, or because it´s wet and the rubber makes the ideal line slower. I think that decisions are what will make the difference

Also, I wonder how a computer will evaluate when braking a bit late is worth to overtake a car or not. This is standard procedure for human drivers, if you go a bit long but not enough to go out of track so you can keep track position, it´s worth even if you did the corner slower than posible. But if too long, even if keeping on track, the car you just overtook will pass you again. Where´s the limit? Programming those limits will be a key factor.

BTW, to stablish those limits I hope they don´t ask Rosberg

All the things you mention about not the ideal line is experience. This can be programmed in.

A corner, with all the variables is actually, if you break it down, the same as a driver takes it.
- you start braking at the point depending where and how fast you want to be when you turn in, calculating the degradations of track, tires and car from the previous laps, when you overshoot for some reason, you calculate a new ideal line from that point. Etc etc etc.

[FW17]

Not sure if that is the point of Roborace

It is not about a pre-programmed lap but AI doing the lap based on inputs received from the sensors

[Jolle]

Not sure if that is the point of Roborace

It is not about a pre-programmed lap but AI doing the lap based on inputs received from the sensors

Just like a real driver I think you're allowed to show it a map of the circuit, just like you can program/teach it have certain tactics and lines when the track is different. "If grip is below x, move line y to z", just like real drivers.

[Andres125sx]

I think where things are more difficult is in the 'vision' side of things - detecting the edges of the track, 'seeing' if the upcoming corner is an easy curve or 'throw out the anchors'hairpin etc, plus picking the correct racing line & knowing when it is fact quicker to take a 'non-optimal' line for one corner in order to be quicker through a following corner, picking braking points etc.

I don´t think the ´vision´ part will be difficult, they can use tons of sensors for that, and they can be placed higher for better view, unlike driver´s eyes wich must be low in order to lower CoG. Also, proximity sensors are a lot more accurate than a driver evaluating if his front wing will hit the car in front or not.

To me the most difficult part will be programming unusual situations. For a qualy lap I think it must be quite easy to program a drone. Actually Audi already did it. The difficult part will be to evaluate when the ideal line is not the ideal line anymore because there´s a car, or because it´s wet and the rubber makes the ideal line slower. I think that decisions are what will make the difference

Also, I wonder how a computer will evaluate when braking a bit late is worth to overtake a car or not. This is standard procedure for human drivers, if you go a bit long but not enough to go out of track so you can keep track position, it´s worth even if you did the corner slower than posible. But if too long, even if keeping on track, the car you just overtook will pass you again. Where´s the limit? Programming those limits will be a key factor.

BTW, to stablish those limits I hope they don´t ask Rosberg

All the things you mention about not the ideal line is experience. This can be programmed in.

A corner, with all the variables is actually, if you break it down, the same as a driver takes it.
- you start braking at the point depending where and how fast you want to be when you turn in, calculating the degradations of track, tires and car from the previous laps, when you overshoot for some reason, you calculate a new ideal line from that point. Etc etc etc.

And when there´s a car in front of you?

And when there´s a car in front of you who did brake too late?



I agree with you this is all experience with programming, but drivers usually have a lot of years to develop theirselves and learn from that experience, while roborace engineers will have to do it all at once.

I expect big differences at the beginning of the championship

[Jolle]

I don´t think the ´vision´ part will be difficult, they can use tons of sensors for that, and they can be placed higher for better view, unlike driver´s eyes wich must be low in order to lower CoG. Also, proximity sensors are a lot more accurate than a driver evaluating if his front wing will hit the car in front or not.

To me the most difficult part will be programming unusual situations. For a qualy lap I think it must be quite easy to program a drone. Actually Audi already did it. The difficult part will be to evaluate when the ideal line is not the ideal line anymore because there´s a car, or because it´s wet and the rubber makes the ideal line slower. I think that decisions are what will make the difference

Also, I wonder how a computer will evaluate when braking a bit late is worth to overtake a car or not. This is standard procedure for human drivers, if you go a bit long but not enough to go out of track so you can keep track position, it´s worth even if you did the corner slower than posible. But if too long, even if keeping on track, the car you just overtook will pass you again. Where´s the limit? Programming those limits will be a key factor.

BTW, to stablish those limits I hope they don´t ask Rosberg

All the things you mention about not the ideal line is experience. This can be programmed in.

A corner, with all the variables is actually, if you break it down, the same as a driver takes it.
- you start braking at the point depending where and how fast you want to be when you turn in, calculating the degradations of track, tires and car from the previous laps, when you overshoot for some reason, you calculate a new ideal line from that point. Etc etc etc.

And when there´s a car in front of you?

And when there´s a car in front of you who did brake too late?



I agree with you this is all experience with programming, but drivers usually have a lot of years to develop theirselves and learn from that experience, while roborace engineers will have to do it all at once.

I expect big differences at the beginning of the championship

The first few outings probably look like toddlers in karts indeed, but because the bots can calculate more variables, have eyes in the back of their heads and possibilities beyond humans (like regulating brake pressures per wheel, less reaction time), within a few years they might be faster then the current drivers.

[FW17]

I thought they would have integrated the wishbones within those structures

[SR71]

https://www.youtube.com/watch?v=nCOWvTlL66k

I thought they would have integrated the wishbones within those structures

BTW, original concept renderings had exposed wishbones. Knowing Daniel Simon was involved I knew the concept renderings wouldn't be far off base from production but this new video shows a car extremely true to design vision.

To me this says there is a lot of money involved and platform R&D was paid for before the project was announced to the press.

Strong evidence of this series having at least a 1st year.

[i70q7m7ghw]

Surely there are huge benefits to the development of driverless cars from this serious? I hope it does go ahead, it's an excellent idea.