A more realistic route to autonomous driving • Hacker News Board

Re: A more realistic route to autonomous driving

#81

Post by bumholio » Mon, Aug 06, 2018, 9:03 PM UTC

pastage wrote: post #70

ghaff wrote: post #59

Earlier quoted context omitted.

The IRS allows a 50 something cents per mile business expense deduction. And if you run some numbers that’s a pretty good ballpark for the cost of operating a car.

Considering they think Waymo might get prices down to 65-95 cents per mile with fully matured tech that seems like good number.

Considering that's more than a cab costs in some European countries, driver included, I would say we aren't talking about commodity tech, but a highly profitable initial phase.

Re: A more realistic route to autonomous driving

#82

Post by mmt » Mon, Aug 06, 2018, 9:23 PM UTC

perl4ever wrote: post #79

mmt wrote: post #77

Earlier quoted context omitted.

> Auto cars can also drive away from expensive parking when not in use. Driving away is, of course, not standing still. That increases the cost, but, perhaps more importantly, increased latency (or, rather, jitter). They need not be stored in the most expensive, densest areas, but, for those who want to get rid of car ownership, that non-zero cost (especially if borne by the public with free on-street parking) isn't…

Depreciation and utilization rates need to be incorporated. If interest + depreciation is about $9K/yr, that's $25 a day, which means if you drive 8 hrs a day and have a 58% utilization rate, is $5.38/hr which is not all that far from what Uber drivers net. At 30 mph, it's 18 cents added per mile. A self-driving car could perhaps operate more than 8 hours a day, but that doesn't mean that customers will be equally av…

> Depreciation and utilization rates need to be incorporated.

That's tough to do, since we don't know either one. I suspect there's also a (perhaps incorrect) assumption that an unused auto-car will be "just sitting there" rather than moving, in which case utilization is, effectively, 100%.

That's not possible with a human driver in at least the case of a long one-way trip. However, with auto-cars, this could end up merely increasing the capital cost at the expense of unpaid/empty trip cost.

The extra cost is something I mentioned in my latency/jitter comment, wrt cheaper parking.

> $5.38/hr which is not all that far from what Uber drivers net

I didn't check your arithmetic, but, if so, it's is quite far from the fair/livable wage of $15/hr that at least I've been assuming for this conversation.

The fact that rideshare services are currently, effectively, paying much less is certainly an indication that even that may not be enough to lure people away from ownership.

Re: A more realistic route to autonomous driving

#83

Post by perl4ever » Tue, Aug 07, 2018, 3:33 AM UTC

mmt wrote: post #82

perl4ever wrote: post #79

Earlier quoted context omitted.

Depreciation and utilization rates need to be incorporated. If interest + depreciation is about $9K/yr, that's $25 a day, which means if you drive 8 hrs a day and have a 58% utilization rate, is $5.38/hr which is not all that far from what Uber drivers net. At 30 mph, it's 18 cents added per mile. A self-driving car could perhaps operate more than 8 hours a day, but that doesn't mean that customers will be equally av…

> Depreciation and utilization rates need to be incorporated. That's tough to do, since we don't know either one. I suspect there's also a (perhaps incorrect) assumption that an unused auto-car will be "just sitting there" rather than moving, in which case utilization is, effectively, 100%. That's not possible with a human driver in at least the case of a long one-way trip. However, with auto-cars, this could end up…

Utilization rate means percentage of time that is billed to a customer, I believe/presume. Whether a car is moving is not the same thing. The issue is whether it is on the clock or not, and given that computers already are used to match passengers and route cars, I'm not clear on why we would expect a drastic increase. Currently, we see 50-60%. Even a perfect routing algorithm can't make it 100%, because there are only so many customers, so things like how many cars there are per customer matter. If you reduce the cars per customer, service suffers. If you increase the number, utilization falls.

Re: A more realistic route to autonomous driving

#84

Post by mmt » Tue, Aug 07, 2018, 6:22 AM UTC

perl4ever wrote: post #83

mmt wrote: post #82

Earlier quoted context omitted.

> Depreciation and utilization rates need to be incorporated. That's tough to do, since we don't know either one. I suspect there's also a (perhaps incorrect) assumption that an unused auto-car will be "just sitting there" rather than moving, in which case utilization is, effectively, 100%. That's not possible with a human driver in at least the case of a long one-way trip. However, with auto-cars, this could end up…

Utilization rate means percentage of time that is billed to a customer, I believe/presume. Whether a car is moving is not the same thing. The issue is whether it is on the clock or not, and given that computers already are used to match passengers and route cars, I'm not clear on why we would expect a drastic increase. Currently, we see 50-60%. Even a perfect routing algorithm can't make it 100%, because there are on…

> given that computers already are used to match passengers and route cars, I'm not clear on why we would expect a drastic increase

It depends on what the denominator is.

If it's only the time between when a driver/auto-car "accepts" the ride and the end of the ride, there's little reason [1] to expect a drastic increase.

However, if the denominator is the total time the driver is "on duty", which is, I believe, what is generally used to calculate rideshare drivers' effective hourly compensation, then my original point stands. That is, an auto-car can be "on duty" even while just sitting in storage.

The current algorithm also doesn't tell rideshare drivers where to be while on duty, only routing them once a ride is requested. In the auto-car scenario, the computer has complete control, so a predictive algorithm could increase utilization, even if the denominator is time-in-motion.

Whether any increase would be drastic is debatable, but there's opportunity for something.

[1] Currently, the computer routing algorithm has an incentive to optimize for time at the expense of distance (since it's the driver who bears the expense of the unbilled distance, AFAIK). In the case of an auto-car, that perverse incentive would be absent, but I don't expect the difference to be huge.

Re: A more realistic route to autonomous driving

#85

Post by zamfi » Tue, Aug 07, 2018, 7:19 AM UTC

perl4ever wrote: post #80

zamfi wrote: post #76

Earlier quoted context omitted.

Of course you're right; my comment was mostly tongue-in-cheek. Actual train operator salaries probably cost BART about $50m annually, not $500m. In both the cases of BART and private cars, the money to pay for tracks or roads comes from taxpayers. The trains themselves are now 40+ years old, and are finally being replaced at a cost of ~$300m this year. Grandparent post specifically talked about operating costs though…

Utilization rates are 50-58% for Uber or Lyft. When you take it into account that a taxi is not en route 100% of the time, and add the cost (interest, depreciation) of a much more expensive vehicle, I think your 65% savings go up in smoke.

Sorry, I'm not smart enough for this comment to stick with me. Could you walk me through how 50% utilization means the savings go up in smoke?

The most compelling thing I can pull from your comment is "autonomous cars will be 3x as expensive as regular cars (at least), so you won't save anything". That may be short-term true, but is almost certainly long-term false.

A 50% utilization compared with 100% utilization makes almost no difference in vehicle cost structure -- if anything, it actually increases the labor fraction of taxi cost.

If you're driving your car 50% of the time, then wear and milage is the dominant depreciation factor (rather than age). But a human sitting around doing nothing but waiting for that 50% of the time doubles the labor (time) cost of the service, suggesting an even greater savings from eliminating the human driver.

What am I missing?

Re: A more realistic route to autonomous driving

#86

Post by Eridrus » Tue, Aug 07, 2018, 5:01 PM UTC

IshKebab wrote: post #50

Eridrus wrote: post #5

This is certainly more cautious, but it's not clear that this is any more realistic than Waymo's plan which involves limiting vehicles to a predefine area, mapping in great detail, etc.

It's clearly more realistic. They don't have to deal with roadworks, they have preset pickup points (no need to figure out finding a safe space to stop), they are relying on other drivers being cautious around a bright orange vehicle rather than trying to act like a real car, they have explicit screens for communication with pedestrians rather than trying to do it implicitly like we would (I.e. slow down to let a car…

Have they significantly reduced the challenge though? I don't think the core task of driving is significantly easier than these problems that they have elided.

I know it is a bit of a fallacy to say "we can fly to the moon, so we should be able to have a self-driving car", but it's also not clear that the tools we have (ML, simulators, planning algorithms, etc) are not up to the tasks that have Drive.ai have engineered around.

It's a reasonable strategy to elide things you can and focus on the "core" hard task, but if solving these other difficult problems is also meaningful product-wise (and it clearly is meaningful), there's no reason you can't have people working on both. And given Waymo seems to be gearing up for a large launch, I have a hard time believing that these additional tasks are as hard as Drive.ai want to make out.

Re: A more realistic route to autonomous driving

#87

Post by perl4ever » Tue, Aug 07, 2018, 8:19 PM UTC

mmt wrote: post #84

perl4ever wrote: post #83

Earlier quoted context omitted.

Utilization rate means percentage of time that is billed to a customer, I believe/presume. Whether a car is moving is not the same thing. The issue is whether it is on the clock or not, and given that computers already are used to match passengers and route cars, I'm not clear on why we would expect a drastic increase. Currently, we see 50-60%. Even a perfect routing algorithm can't make it 100%, because there are on…

> given that computers already are used to match passengers and route cars, I'm not clear on why we would expect a drastic increase It depends on what the denominator is. If it's only the time between when a driver/auto-car "accepts" the ride and the end of the ride, there's little reason [1] to expect a drastic increase. However, if the denominator is the total time the driver is "on duty", which is, I believe, what…

"That is, an auto-car can be "on duty" even while just sitting in storage."

If you consider the robot to "work" for more hours than a human, that's great, but due to the lower average revenue, it needs to be cheaper in order to be competitive with humans. There's no way to move your self-driving car to the opposite side of the world for the night.

"Whether any increase would be drastic is debatable, but there's opportunity for something."

If self-driving cars are cheaper, it seems like that would lead to more of them driving longer hours than humans, which would lower the utilization rather than increase it.

Re: A more realistic route to autonomous driving

#88

Post by perl4ever » Tue, Aug 07, 2018, 8:23 PM UTC

zamfi wrote: post #85

perl4ever wrote: post #80

Earlier quoted context omitted.

Utilization rates are 50-58% for Uber or Lyft. When you take it into account that a taxi is not en route 100% of the time, and add the cost (interest, depreciation) of a much more expensive vehicle, I think your 65% savings go up in smoke.

Sorry, I'm not smart enough for this comment to stick with me. Could you walk me through how 50% utilization means the savings go up in smoke? The most compelling thing I can pull from your comment is "autonomous cars will be 3x as expensive as regular cars (at least), so you won't save anything". That may be short-term true, but is almost certainly long-term false. A 50% utilization compared with 100% utilization ma…

I was assuming humans and self-driving cars both get revenue from riders, but don't get paid in between. And I'm interpreting utilization as the percentage of the time that riders are paying.

Re: A more realistic route to autonomous driving

#89

Post by mmt » Tue, Aug 07, 2018, 8:51 PM UTC

perl4ever wrote: post #87

mmt wrote: post #84

Earlier quoted context omitted.

> given that computers already are used to match passengers and route cars, I'm not clear on why we would expect a drastic increase It depends on what the denominator is. If it's only the time between when a driver/auto-car "accepts" the ride and the end of the ride, there's little reason [1] to expect a drastic increase. However, if the denominator is the total time the driver is "on duty", which is, I believe, what…

"That is, an auto-car can be "on duty" even while just sitting in storage." If you consider the robot to "work" for more hours than a human, that's great, but due to the lower average revenue, it needs to be cheaper in order to be competitive with humans. There's no way to move your self-driving car to the opposite side of the world for the night. "Whether any increase would be drastic is debatable, but there's oppor…

> due to the lower average revenue, it needs to be cheaper

It's unclear to me, but this may be tautological, or at least reversible (by being cheaper to be competitive, it reduces revenue). Perhaps I'm missing your point?

> There's no way to move your self-driving car to the opposite side of the world for the night.

Ah, but that's moot. Unlike with a human driver, an auto-car doesn't need to be moved to the opposite side of the world (or an approximation, like Sacramento from SF) for the night.

> If self-driving cars are cheaper, it seems like that would lead to more of them driving longer hours than humans, which would lower the utilization rather than increase it.

You're still confusing "driving" (in motion) with "on duty".

Also, even if an auto-car is cheaper while driving, that doesn't matter if the different parties are bearing the cost of off-meter driving. In the auto-car case, it's the vendor (e.g. Uber), so there's a strong incentive to maximize utilization. In the human case, it's the driver, so the vendor has no such incentive (nor even the ability during "on duty" but not driving-for-that-vendor times).

Re: A more realistic route to autonomous driving

#90

Post by zamfi » Thu, Aug 09, 2018, 4:32 AM UTC

perl4ever wrote: post #88

zamfi wrote: post #85

Earlier quoted context omitted.

Sorry, I'm not smart enough for this comment to stick with me. Could you walk me through how 50% utilization means the savings go up in smoke? The most compelling thing I can pull from your comment is "autonomous cars will be 3x as expensive as regular cars (at least), so you won't save anything". That may be short-term true, but is almost certainly long-term false. A 50% utilization compared with 100% utilization ma…

I was assuming humans and self-driving cars both get revenue from riders, but don't get paid in between. And I'm interpreting utilization as the percentage of the time that riders are paying.

The human's time has a cost, and even if the rider doesn't explicitly pay it, it is factored in to the labor cost of the driver.

The machine's time cost is time-based depreciation, which is pretty low.