Maneuverability for AI Self-Driving Cars


By. Dr. Lance Eliot, the AI Trends Insider

Float like a butterfly, sting like a bee.

You might remember or have heard this famous quote which was uttered by Muhammad Ali in 1964. He was referring to his boxing prowess at the time. He was young (age 22) and a boxer that delighted in being on his toes and weaving and bobbing around the boxing ring. Able to be as agile as a butterfly, he could move so readily and quickly that it was hard for his opponents to land a punch. This clever turn of a phrase has become legendary and used whenever someone wants to claim they are agile and fast.

Driving a car is somewhat the same.

Novice drivers are slow to react and often drive a car as though it is a Sherman tank. You’ve likely sat behind a student driver while on the highway or on a crowded city street. If so, you know how exasperating it can be. The learning driver will start and stop, sometimes for no apparent reason. They will take curves at an agonizingly slow pace, typically an inch at a time. They often hog the road and make it hard to get around them. They are the opposite of a butterfly. Maybe one could say they are more akin to an elephant, lumbering and causing traffic around them to bunch up and slow down.

One especially noticeable aspect of the novice driver is that they tend to pick a lane and stay with it, until they are pretty much forced to change to a different lane. This makes sense in that it is easier to just drive along in a particular lane, and from their viewpoint switching lanes is overly tricky. To switch a lane you need to gauge the traffic in your lane and the lane you want to go into. You are supposed to signal and wait for the right moment to make the switching move. You need to guide the car into the other lane and do so without steering into another car. You need to watch your speed. You need to keep from hitting the car ahead of you. You need to keep from getting hit by another car. And on, and on.

Proficient drivers scoff as this kind of a maneuver in the sense that it is easy for them to pull it off. They cannot imagine why a novice driver has so much trouble changing a lane. For gosh sakes, it’s just a lane change, the proficient driver exclaims. No big deal. But, if you ask a novice driver, they will tell you about the nightmares they have of dreading making lane changes. The need to pay attention to so many variables simultaneously that it is mentally overloading. The danger factor is extremely high and they realize that one wrong move and it could be a bad day for them and everyone else involved. I’ve seen some novice drivers that never reach a sufficient proficiency level and pretty much stay off the highway because they want to avoid fast-speed lane changing.

Maneuvering a car is easy to speak of, but doing it successfully requires building up an appropriate skill.

Go back in time to when you first drove a car, and I’m betting you’ll be able to remember the cold sweats when you were first sat behind the wheel of a car. It’s a huge responsibility. You will determine the fate of the car, your own life is now in your hands, and the lives of any occupants in your car are in your hands, as are the drivers and occupants of other cars that you’ll come across while driving. Pedestrians lives are in your hands. Dogs and cats lives are in your hands. In whatever manner you decide to maneuver the car, you can either hit, injure, or kill living things. Most of us drive our cars every day and don’t think overtly about any of this. Instead, we just take it for granted that we have in our hands the destiny of those around us.

Over time, a savvy novice driver should be improving in the skills at maneuvering a car. They gradually learn how to go more quickly around curves. They progressively get more comfortable switching lanes. They adjust mentally and learn new tricks of driving. Driving becomes less of a conscious skill and begins to come naturally. The car becomes an extension of their own arms and legs. There is a flow of sorts that happens and the car and the driver become one. I know this sounds like some kind of zen thing, but I am just saying that the novice driver becomes more comfortable in the driving role and eventually settles into it.

Even proficient drivers though can be shocked back into the reality of how difficult it is to maneuver a car successfully. A colleague of mine that drives a nice sports car is continually bragging about how he can zoom around in traffic. He tells tall tales about how he was able to get to work in record time by “scaring the sheep” drivers and having them get out of his way (he aims his car at other cars, and plays a game of chicken to get them out-of-his-way – a dangerous and foolhardy gambit). Well, one day, he got into a car accident, which fortunately no one was seriously hurt, but he was taken aback and suddenly remarked that maybe he was cutting corners a bit too much. The NASCAR cockiness for a moment had receded and he had a realization that cars are large physical objects with a potential for harming him and others.  He had forgotten this.

It turns out that maneuverability is a topic that has been studied closely by animal behaviorists. A recent interesting study examined how hummingbirds are able to be so nimble. In the study, computer vision systems were used to record and analyze the movements of more than 200 hummingbirds. We are all familiar with how agile hummingbirds are, and they seem to shift direction on a dime. What is the morphology that explains this capability? The analysis showed that it is a combination of the shape of their wings and as combined with their skills. These marvelous creatures are able to rapidly fuse together sensory data about their environment, and use it to then avoid obstacles and do so at incredibly breakneck speeds.

The study included 25 different species of hummingbirds. Certain of the species would use one kind of maneuvers, while a different species would choose to use a different set. Aspects such as forward acceleration and deceleration were largely determined by muscular capacity. Wing shape and size made it either easy, or hard, for making particular kinds of maneuvers. Diversity in their ability to fly raises questions of why not just have the same capability for all hummingbirds. It is thought that evolutionary factors ultimately led to different variants, and each distinct set of factors then leads to a hummingbird that is best suited for its environmental niche.

When you translate this study of hummingbirds over to the domain of cars, you can consider that the physical aspect of the car determines some of the maneuverability aspects, akin to how the shape and size of a hummingbird wing determines the limits of what maneuvers can be made. With hummingbirds, it’s more than just their physical shape, since they also need to perceive their environment and their brains need to engage as to how their bodies are to react. Likewise, when driving a car, the human driver needs to mentally gather sensory information, fuse it together, and cognitively make decisions in real-time to control the car.

For a car, the size of the car is one determiner of maneuverability. Think about a subcompact sized car versus the maneuverability of a giant sized SUV. How fast does the car go from zero to 60 miles per hour? What kind of tires does the car have, just normal traction or high-performance tires? There are numerous physical aspects about a car that determine what kind of maneuverability it can potentially have. In the end, it is the mind and hands of the driver that will showcase the physical maneuverability of the car.  A novice driver tends to drive all cars the same way, and has not yet learned how to leverage the potential maneuverability of the car.  A proficient driver is able to ascertain what the car can do, and has the capability to make use of the car’s proficiencies.

The environment surrounding a car will also significantly shape the maneuverability options. If it is raining and the roads are slick, a car that could normally do the zero to 60 miles per hour in 4.4 seconds on a dry road is unlikely to be able to do so on a wet road, even if the human driver is highly proficient. The physics of the road conditions and the car will provide limits on what the human driver can do. Besides the weather and the roadway, there’s also the factor of traffic. If you are in a superfast car that has incredible maneuverability, and suppose you are the top car racer in the world, but if you are on the Los Angeles 405 freeway at 5:00 p.m. on a weekday, when the freeway is completely clogged with cars and its stop-and-go at less than 5 miles per hour, your keen maneuverability skills and car capabilities will be stymied.

Allow me to add another factor into driving maneuverability. I mentioned earlier about my colleague that bragged about his driving prowess and used bulldog tactics to get his way. I’m going to suggest that we consider driving maneuverability to include something that he does not exhibit, namely being able to drive with grace and aplomb. I say that because mindless darting into and out of lanes does not make for a highly successful sense of maneuverability. Any clod can do that. A truly maneuverable-capable driver is able to maneuver in such a manner that they have a neutral or even positive impact on the rest of the traffic around them.

Say what? I am sure some of you are perplexed by my claim. Allow me to further elaborate. If you are maneuvering a car in traffic and able to get where you want to go, efficiently and effectively, and simultaneously avoid negatively impacting other traffic, you are doing maneuverability the way it is intended to properly be done.  In contrast, the sloppy and careless maneuverable driver is one that makes lane changes at the last moment and forces other cars to jam on their brakes. These low-skilled maneuvers are done in a fashion that increases the risks and danger of driving for them and the other drivers around them.  A successful driver is one that has high maneuverability skills, and for which they combine those skills with the nature of the maneuverability of the car and the maneuverability allowed by the nature of the roadway environmental conditions.

What does this have to do with AI self-driving cars?

At the Cybernetic Self-Driving Car Institute, we are developing AI that enables self-driving cars to drive with a high degree of successful maneuverability.  They are like hummingbirds.

You might be wondering, well, Lance, won’t all self-driving cars be designed and developed with AI for a high degree of successful maneuverability? The answer, right now, is no. Currently, most of the auto makers and tech firms are zipping ahead with getting their self-driving cars on the roadways and the “clod” approach to driving is just fine, they believe.

If you observe how today’s AI self-driving cars maneuver, you would realize quickly that they are driving like a novice. They get into a lane, and hesitate to get out of the lane. When they switch lanes, it’s done only in the safest of circumstances and the self-driving car acts very timidly. The maneuverability factor is low, in terms of the driver. It’s not the car, since the cars are generally the same kinds of cars that we humans drive. It’s the AI that’s not up to par.

We want to have AI that can drive a car as a proficient human would drive a car. A Level 5 self-driving car is one that can be driven by the AI and does not need any human intervention. The self-driving car industry is not yet at a Level 5 self-driving car, nobody has achieved it. Even once we get there, the Level 5 criteria does not explicitly say how good a driver the AI must be. In other words, you could have a clod of an AI driver. But, is that what we all in our hearts want a Level 5 to truly be?

AI self-driving cars are going to be mixing with human driven cars for quite some time. Let’s suppose that the first round of AI self-driving cars at the Level 5 are all clod drivers. In that sense, none of them are very good at maneuverability. When they make driving maneuvers, it is stilted and tends to adversely impact the rest of traffic. This is going to be a dangerous situation. Imagine that we have say 50 human driven cars on a stretch of freeway, and we mix into it 10 of the AI self-driving cars that are driving like novices. The odds are that you’re going to have human driven cars that will have a difficult time dealing with this many novices all at once.

We are used to having a few student drivers mixed into traffic, but not have them be say 10% of the traffic, or 30% of the traffic, or half of the traffic. It’s a recipe for disaster. I realize the utopian view of self-driving cars is that the world will consist of only self-driving cars. This is nonsense. We cannot overnight replace 200 million conventional cars (in the United States) and suddenly have all self-driving cars. Some have said that we should have special lanes on our roadways that are dedicated to self-driving cars, and not allow human driven cars to mix in with them. This has a steep price in terms of altering our roadways to do this, plus there will be angry debates about whether this is allowing the privileged (those able to afford self-driving cars) to have something that the rest of the populace cannot readily have.

Let me repeat my earlier statement about driving: A truly maneuverable-capable driver is able to maneuver in such a manner that they have a neutral or even positive impact on the rest of the traffic around them. Now, let’s insert into the sentence that the driver is AI.

Thus, here’s the modern version:

A truly maneuverable-capable AI self-driving car is able to maneuver in such a manner that it will have a neutral or even positive impact on the rest of the traffic around them.

It’s not enough to just make a maneuver. The maneuver not only needs to be safe, but also must be done with a perspective on the impact to the rest of the traffic. The desired impact is either neutral or positive. The undesirable impact is one that is negative.

Here’s a traffic example to illustrate.

Imagine you are in a lane of traffic. The car is going at 50 miles per hour. You want to change lanes into the lane to your right. There is an available gap in traffic in that lane. You can make the lane change, legally, and without hitting another car. Seems easy. Just go ahead and change lanes.

Suppose that I tell you that in the lane that’s to your right, there’s a car coming up in that lane. It’s not yet reached being parallel to your car, and so there is still a physical gap for you to move into the lane. That car though is moving at 65 miles per hour. It is rapidly closing the open gap. If you move into the lane, it will force that other driver to hit their brakes to the degree that they need to rapidly decelerate their car.

I think we would all agree that any driver that jumped into another lane, and forced an upcoming driver to have to rapidly hit their brakes, would be taking a dicey maneuver. Yes, you could say that the other driver should be watching for a car that jumps into their lane. Yes, you could say that the other driver should let in that car. But, I think it’s relatively evident that if the first car caused the second car to hit their brakes, and suppose that behind that second car there was another third car that now had to hit its brakes, the first car has started a cascade of negative impacts on fellow traffic.

Most of today’s AI self-driving cars are driving on the roadways and playing a chess game whereby they only are looking at their most immediate move. Can you move the rook into that open space on the chessboard? Yes, but then the next move after that will allow the queen to take your king. AI self-driving cars are not yet looking ahead of their next move and figuring out the consequences of the moves. They are one ply (meaning looking at a single move), rather than say three ply or five ply (looking ahead by three or five moves).

The AI needs to be using the sensor data to anticipate the adverse impacts on traffic due to the maneuvers potentially to be made by the self-driving car. As part of the sensor fusion and virtual model updating, the AI needs to be planning for what will happen if any given move X is going to be attempted. Can the AI self-driving car maneuver into the lane to the right, and if it does so, will it cut off other traffic? Will it create a cascading set of car collisions or issues? This is what must be done to be a proficient driver of a self-driving car.

For some auto makers and tech firms, they are using large datasets of driving data to train the AI via the use of neural networks to be able to find patterns of driving and act accordingly. This can be handy, but it can also be narrow in that it only is examining the data being provided. If you fed it data of cars changing lanes, and you did not include the consequent traffic repercussions, you would essentially be training the AI to be a clod driver. It is crucial that the datasets of driving data be sufficiently robust to enable the neural network to be aware of a large enough picture of the roadway and road conditions.

Ultimately, we want the AI to float like a butterfly, and not sting like a bee. Efforts toward making AI that is a proficient driver, one with a degree of successful maneuverability, must be a key goal for all self-driving car makers.

This content is originally posted on AI Trends.