Figuring out parking for robocars


People are working hard to get robocars to handle public streets, but they also need to handle private parking lots for parking, pick-up and drop-off. Private lots have all sorts of strange rules, so a system is needed to make it easy to map them and make those maps and rules available to cars. I outline such a system in a new Forbes site article found here:

How Self Driving Cars will figure out Parking


Parking and driving are two fairly different problems. Parking is fairly simple, I think. Either a human (in the car or in a remote location) okays the parking location or the parking locations/times/rules get okayed by the property owner ahead of time. For most places right now it's a simple matter of selecting which parking stalls are allowed to be used (and possibly what times use is allowed). Further into the future that'll change, but by then there will be lots of parking space dedicated solely to serving robocars, and that's a fairly easy problem to solve.

Driving in a private parking lot is a much more difficult problem; but I'd argue that it is a problem that has to be solved in order to have successfully created a robocar for public roads. Sure, you can fake it by relying on human-curated HD maps or on accurate lane markings and other traffic control devices, but the real world is too messy for that. This is especially true in private parking lots, but it's also true on a great number of public roadways as well.

A robocar needs to be able to figure out how to drive when there are no lane markings, no maps, and essentially no rules (difficult private lots don't have strange rules, they have what is essentially no rules). It should be doable, but I haven't seen any evidence that anyone has even come close to building a car that can do it. Hopefully I'm just not up to date with the state of the art.

My feeling is, why does it have to park with no maps? The owner of the lot either wants robocars (including those of the owner, employees or customers) to park there or doesn't. If, for a fairly modest fee, they can get a map of the lot made, why not? There are two functions -- parking/waiting, and pick-up/drop-off. For the latter, if the lot does not have marked spots for pick-up/drop-off, you use te closest public spot (which will be mapped by the car operator.)

If a person is with the car when it parks, they can solve any problems of unmapped lots.

My comment about needing to be able to do it with no maps was about driving, not parking. I don't think using the closest public spot to drop passengers off is reasonable. And I don't think you can call your car a robocar if it can't handle driving through the vast majority of parking lots. My comment was definitional. If a car can't figure out how to drive when there are no lane markings, no maps, and essentially no rules, I wouldn't call it a "robocar." "No new infrastructure" is one of the key requirements to be a robocar. Up-to-date, human-curated, detailed maps of every driveable location on earth is infrastructure.

For parking, I agree, although I think in most instances it won't (and shouldn't) even require a fee. Just for the owner to look at a map and click on which parking spots are allowed. In some cases, if parking is tight, the robocar company might even be the party paying the fee. You say "the owner of the lot either wants park there or doesn't." But it's actually not that black and white. The owner may very well want some robocars to park there and others not to. There's going to be competition. A robotaxi company that tries to charge land owners fees in order to park on their land will quickly lose out to other robotaxi companies. That actually goes for drop-off as well, although only the most popular destinations (e.g. stadiums, airports, Disney World) will get away with restricting which robotaxi companies can use their best drop-off locations.

In essence, a robotaxi with passengers in it should be able to go virtually anywhere a taxi can go. That includes being able to drop off those passengers and then proceed safely and relatively courteously to its next destination, whether that be a parking spot or to pick up a new passenger. Not all parking spots need to be fair game, and not even all areas of a parking lot need to be fair game, though practically speaking there likely does need to be parking available somewhere relatively close for all but the most crowded locations. A robotaxi without passengers in it should not drive or park on private land except with permission of the land owner, as necessary to leave an area after dropping someone off, or to pick someone up who has explicitly called it to that location (and the latter only if the land owner does not explicitly prohibit such pick ups).

Also, of course, many cars will have options for someone to take over the wheel in locations that aren't adequately mapped. I don't think you can call such a car a robocar, but it will be extremely useful to have a car/service that can drive autonomously on most public streets and in a few very popular private parking lots, and can drive manually everywhere else.

Later on (or maybe earlier on), I definitely foresee cars with no steering wheels, but which are still operated by humans. A steering wheel is a fairly poor device to provide the sort of human input that not-quite-robocars need. They're mostly an appendage of the days before drive-by-wire. It'll be interesting to see how the law treats such a car.

As you say, it can often be constructed from an ariel photograph. The main concept of the "map" is just that a human being has confirmed the rules of the lot that a robot may have trouble figuring out, because unlike roads, there is not necessarily a pattern to parking lot rules that is universal. And for new concepts, like areas where you should ignore the lines and valet park, there is no system and virtual infrastructure is the cheapest way to do it.

I'm not sure if your comment was about parking or driving through lots or both, but I think that makes a big difference.

For parking, a robotaxi company is probably going to need to get permission from the land owner anyway, so it's not a big deal to make a map while doing that. Additionally, if you can only park in 1/10th of the parking lots in a city, that's more than enough.

But for picking up / dropping off, I just don't see how maps are very helpful. It's true that "there is not necessarily a pattern to parking lot rules that is universal." But the "rules" (in quotes because they're not really rules, but customs), especially in places where they deviate from local customs, are also not particularly human-mappable.

I think this is true with public roads too, by the way, albeit less so, especially in countries with better-enforced driving laws.

Driving to a large extent involves playing follow-the-leader, not following particular rules. This is true sometimes on the public roadways (a good example would be the driving I did this past July 4 trying to find parking for the fireworks show), and it is true more frequently in parking lots. You can't map that kind of stuff. It's ad-hoc, and it changes frequently.

For large venues (stadiums, amusement parks) or major events (4th of July, New Year's Eve), once robocars are common enough, there will be a need for infrastructure. But here I'd disagree with you that physical infrastructure doesn't make sense. In these sorts of situations, where there's a lot of robotraffic, it makes sense to give the robotraffic its own dedicated lanes. And it's not expensive. For a temporary event, it could be as simple as putting out a bunch of traffic cones and a "robocars only" sign. Yeah, they can notify the companies to let them add the entrance/exit to the dedicated lanes to their maps. But they don't need highly detailed maps for that.

Large permanent venues will benefit from having more expensive infrastructure for robocars. But this will be instead of the (even more) expensive infrastructure they have for human driven cars.

Eventually, when/if robocars start taking over the world, even smaller venues (supermarkets, fast food drive-thrus) will probably start having dedicated robotraffic lanes.

Because all it has to be is to draw a box around an area and say, "robots park here under standard robot valet rules #4." Then put up signs saying robots only. And yes, robot valet rules #4 has rules about what to do when humans park there anyway.

Yeah, for showing them where they can park, it's easy to map it.

Give lot owners the option of designating robocar lanes and waiting spots. These can be painted and/or specified by lat/long points. The first mover creates an open standard that other robocars can use. The first mover would do the heavy lifting of tagging the lanes at the lot owner's request.

For lots without designated robolanes just use a feature like Tesla's enhanced summon where the car navigates very slowly toward the rider's phone. Since the rider chooses the pickup spot there's no need for the car to "understand" every parking lot on the planet.

From what I've seen of Tesla's enhanced summon, it is truly awful. No understanding whatsoever of how to drive in a parking lot. Yeah, it can avoid hitting things (usually, except apparently for curbs), but that's about it.

I'm sure some other companies are doing better, and I'm sure Tesla will also get better over time. But it's going to take real AI, not detailed maps and/or hand-programmed rules.

(I was somewhat surprised by how awful Tesla enhanced summon was. The car doesn't seem to even understand how to drive on roads that don't have lines.)

You can figure out many parking lots but because there are no formal rules, a small layer of virtual infrastructure (maps) is much cheaper than new signs or painting lines. Why do it the hard way with physical infrastructure?

I sure hope these stories about $50 lidar are true. That'll make driving through parking lots a lot easier.

Actually, if you just want a sensor with short range 3D sensing, those are already much less than $50, using phased based time of flight, or even 3D ultrasound. The range is typically about 10m which is workable for parking lot speeds. People use them on lower speed robots, they are not found on cars because the cars already have higher ability sensors.

My hopes that $50 lidar sensors become a reality are not limited to the impact that'll have on driving in parking lots.

I imagine Tesla' Enhanced Summons relies on their ultrasound sensors to give them reliable object location info. This satisfies the "Thou shall not hit stuff" commandment even when their image recognition NNs fail.

Of course that just gets you to the starting line. From there you can begin to solve the myriad real problems of self-driving, even if it's just in a parking lot.

They seem to have a problem with curbs. Lidar could probably help, though more cameras would likely help as well. I'm not sure how precise the ultrasound sensors are.

At least for now, summons isn't self-driving. It's remote-control driving, albeit with very poor controls.

Add new comment