By Dr. Lance B. Eliot, the AI Trends Insider
Most of us have at one time or another tried to parallel park our car and suffered the frustration and at times humiliation that we just could not seem to get the car into that desired space. Back and forth, trying over and over, meanwhile sometimes there is someone watching as we do so, such as pedestrians on the sidewalk or maybe a driver sitting behind you that is waiting for you to get out of their way and just finish that darned parallel parking job. It can be nail biting. Plus, though you aren’t willing to likely admit it, I am betting that you probably touched or bumped a car or two while trying to squeeze into a tight parallel parking space. You undoubtedly looked around to see if anyone spotted you doing so, and if not, quickly disembarked from your now parked car and pretended that you had not played bumper cars with strangers.
When you start out as a novice driver, the act of parallel parking seems incredibly difficult. Many teenage drivers that are careful about their driving will purposely avoid doing parallel parking. Those that don’t care about other cars are often eager to do parallel parking, tending to do so at a speed and manner that would make their parents become queasy if they knew that their teen driver was muscling their car this way. Elderly are prone to avoid doing parallel parking, partially due to the complexity, but also at times due to the twisting and contortions that as a driver you often need to do, looking backward and over your shoulder as you execute the maneuver. This can make anyone’s back and neck be sore, no matter what your age.
Does all of this indicate that parallel parking is a highly complex task? No. In fact, it is a relatively straightforward task. At least when considering the act of carrying out the formulaic parts of the parallel parking task. In a moment, we’ll walk through the tasks. As will be shown, this is pretty much a mindless kind of driving task. From an AI perspective, there’s not much involved. One might be cynical and say that you could train a monkey to parallel park. By this, I mean that the steps are routine, easily trained, repeatable, and requires almost no human judgment per se.
At the Cybernetics Self-Driving Car Institute, we are aiming at stepping up the parallel parking task in ways that do require human judgment and provide AI specialized software that goes far beyond today’s mindless AI used for parallel parking. If you enlarge your viewpoint of the scope of the parallel parking task, you’ll see that there are aspects today that are being omitted entirely by self-driving cars that have a so-called parallel parking capability, often called Auto Park or Active Park Assist.
First, what are the steps required to parallel park a car? We’ll make our scope narrow and then enlarge it after we’ve covered the basics of this all-to-common driving task.
Here’s what you are normally urged to do:
- Pull alongside the lead parked car, allowing a 3-foot gap parallel between you and it
- Align your back tires with the lead parked car’s rear bumper, halt
- Go into reverse and before starting motion turn your wheels hard-right
- Start into motion, backing up slowly, until you are at an angle of 45 degrees, halt
- Now, while not in motion, turn your wheels hard-left
- Continue slowly to back-up, until you are in the spot and parallel with the curb
- If needed, move backward or forward to even out the space ahead and behind you
There are other ways to do parallel parking, but the above is well representative of the recommended approaches. If you’ve not ever had someone lay out the step-by-step approach, well, guess what, you now know formally how to do parallel parking. Many people were never schooled in how to do so. They learned by trial-and-error. Indeed, some never figured it out and so just chock up the whole thing to a magical method that they can’t seem to divine.
The truth be told, there is no magic involved. In the self-driving car industry, you find some that smirk at consumers that marvel at a car that is able to parallel park by itself. What magic! What a wondrous sight to see! The human driver takes their hands off the wheel, and with seemingly a robotic ghost driving the car, it smoothly works its way into a parallel parking space. This must be AI at its finest. Maybe this is a sign that we are soon on the verge of AI taking over the world. We are soon to be slaves of the AI systems that will rule us.
It’s a simple parlor trick. Just like watching a magician that makes a coin disappear and magically reappear, once you know the secret of the trick, you realize that there’s not much to it.
In this case, the self-driving car uses it sensors to detect that the car is in the right placement to start the parallel parking task. It can use the camera, LIDAR (see my column about LIDAR), or radar to measure the distance and identify the parked cars that are nearby. It is connected to the car’s controls such that it can accelerate and brake as needed, along with being able to turn the steering wheel. Following the above formulaic steps, it carries out those steps. One step, then another. Programmed easily. Indeed, it can do this in an exacting manner and without any emotion, which kind of helps. Human drivers sometimes doubt the approach and misbelieve that they are about to hit the parked car, when in fact there are inches to spare. The radar, camera, and LIDAR are not emotional about this and can make relatively accurate measurements and proceed in a coldly calculated manner.
Human drivers are rushed by another car that has come upon them and are waiting anxiously for you to get out of their way. Or, with pedestrians watching you, you are fearful they are going to record your actions and post it onto YouTube, wherein a million hits will arise of you fumbling to make that parallel parking job. Maybe you reverse too fast, or maybe you don’t turn the car wheel far enough. There are lots of ways that a sloppy or nervous human can mess up this procedure. Automation doesn’t get distracted by those aspects and in a rote manner carries out what it was programmed to do. I’ve somewhat discussed this in another piece that I wrote about self-driving cars and parking, so you might want to take a look at that column.
The ability to parallel park is gradually becoming a standard feature on most cars. The automation for it is easy to put into place, once you have the sensors on the car and the automation connected to the controls of the car. The software to do the parallel parking task is easy. It’s having the sensors and the controls to drive the car that are “harder” merely because of the cost of the hardware and its needing to be integrated into the car.
In fact, we now have car makers that are crowing about how their parallel parking feature is faster than someone else’s. One car maker says their car can parallel park in 1 minute. Another one says it can be done in 30 seconds. Now, we are inching our way down, and the latest benchmark seems to be 24 seconds. This is though really kind of not interesting from an AI viewpoint. Shaving a few seconds off of parallel parking is merely more of the same parlor trick. You can control how fast the car reverses and tweak the formulaic approach, but it is still the same mindless task.
Consumers were surveyed and asked whether they trust their car to do parallel parking (assuming that such a feature is available on their car). Interestingly, only about 25% said they would trust a car to parallel park. If the task is as routine as I claim, why aren’t more consumers convinced of the trustworthiness of the parallel parking capability of modern cars?
It could be that consumers are unsure of the formulaic aspects and believe that the task is complex, thus, they assume that the mental abilities of the automation must be quite high to be able to do the task. This makes sense since humans themselves find the task to be bedeviling, and so this carries over into their assumptions about what the automation has to do. Or, it could be that they have rarely seen or used the feature, and so the unknown aspects of whether it really works or not is causing consumers to be hesitant to believe that it works. Maybe after seeing the task done over and over by self-driving cars, humans will eventually believe that it is doable by the automation.
Another explanation has to do with the scope of the parallel parking automation. And, this goes toward my point earlier about how the scope is so constrained that we have taken the judgement out of the equation. Remember that I said earlier that this task is repeatable, simplistic, formulaic, and requires no judgment. Let’s enlarge the scope and see if those assumptions continue to hold true.
First, what happens when you initially even conceive of the idea that you might want to parallel park your car? You have reached a juncture where you want to park your car, and presumably there are no other viable choices other than doing a parallel parking action. In other words, you would have looked around, wherever you are, and determine that the only available parking spot, or at least the only most desirable spot, will require you to invoke parallel parking.
There aren’t any self-driving cars today that scan the surrounding area to try and detect where you can park your car, and then if identifying that the only viable parking spot requires parallel parking, opt to go into a parallel parking mode. Instead, you, the human driver of the car, need to first find such a spot. Furthermore, you, the human driver, need to drive the car to the spot and position the car into kind of just the right posture for it to then become engaged to do the parallel parking.
This is like having a child that can do only one thing and it is up to you as an adult to guide the child to the precise place that allows the child to perform their one magic trick. Unfortunately, this is not going to be sufficient for Level 5 self-driving cars, in the sense that a Level 5 is supposed to be able to do anything a human driver could do (see my column about the Richter scale of self-driving cars), and so requiring a human occupant to have to identify and then position the self-driving car is not going to work. We’ll need the AI to do this.
Next, even if the car is positioned in the right place, today’s parallel parking approaches do not take a look at see if the spot is a legally allowed parking space. Suppose there is a sufficient gap to park your car, but the curb is painted red. Today’s self-driving cars don’t realize this. What about a white painted curb, or a blue painted curb, or yellow, or green? We would want the AI to figure out whether this is a legal spot to park in. This also includes looking at the road signs (see my column on self-driving cars and road signs). There might be a road sign that says no parking in that spot between the hours of 9 a.m. and 5 p.m., and if it is currently noon then you might be inviting yourself into getting a parking ticket. Perhaps there’s a fire hydrant there and so you can’t park there. Etc.
Alright, let’s consider the next aspect, namely, suppose that the spot is blocking a driveway or the sidewalk. This is often a legal issue. Plus, even if in some cases legally allowed, there are some places that parking just doesn’t make sense because of the ire of others that park nearby. I’ve seen angry neighbors that had their driveway partially blocked on a busy July 4th weekend that then decided to run a straight edge along the pretty paint job on the outside of the car. I am not advocating taking the law into your hands, and just emphasizing that there are some spots that might seem viable due to the size of the spot, but there are lots of other considerations that need to be taken into account too.
Identifying the spot to park the car is usually done by simply measuring the gap between the two already parked cars (or, whatever else is serving as the bookends). A typical parallel parking automation will gladly try to fit into a sufficient sized spot, but suppose that someone has parked their bicycle in that same spot. Most of the existing systems don’t look for objects that might be in the spot already. There could be a person standing in the spot and trying to reserve it for their friend. Or suppose a child is standing in the spot and looking for their penny that they dropped onto the street near the curb. These obstructions are rarely looked for, and often if scanned for are only found once the self-driving car is already partially into the self-parking operation.
The only criteria being used currently by the automation involves adhering strictly to the formulaic steps, and then aiming to park as close to the curb as feasible, often ending up at a mere half inch away from the curb, and that the automation wants to keep to a minimum the number of maneuvers taken and the number of curb strikes encountered. This criteria is though not very human-like in many respects.
For example, parking a car within a half inch of the curb can be a bad thing to do. When you open the car door, it might brush against the curb. Or, the car door might swing open and hit a standing object on the sidewalk that is adjacent to the parking spot. Humans usually leave about 8 inches, partially by being lousy at parallel parking but also at times because it is more convenient when getting out of and into the car (and there are laws about how far away from the curb you can legally park, plus, how you turn the wheel depending upon whether you are parked on an incline). There are other instances where you purposely do park close to the curb, such as to avoid stepping into a puddle at the curb. These are aspects that a human driver would consider, and for which we should expect any robust AI to also consider.
Let’s revisit the earlier point about the aspect that surveys of consumers seem to indicate that they don’t trust a self-driving car to do parallel parking. This makes a lot of sense in that the existing parallel parking is being done essentially by a child. The automation is working in a routinized fashion and has no real smarts to it. Why would you trust this? Also, you, the human, need to do so much upfront work to get the car to the proper spot and make sure that it is clear and legal, you might as well finish the job yourself.
I’ve seen some car owners that show off the parallel parking automation to their friends, and then they opt to never use it again. It is one of those boastful features, but not a truly practical day-to-day feature. Of course, the proud car maker can brag about the feature and advertise it. This works to attract buyers of the car. Once the car is bought, the owner though realizes that it is another one of those fancy features that does not cut the mustard. There it then resides, unused. All because it is insufficient. At the Cybernetics Self-Driving Car Institute, we are developing an advanced parallel parking AI component that will make parallel parking into a smart capability and one that consumers will have faith that it can do the job. Drive safely out there.
This content is original to AI Trends.