By Dr. Lance B. Eliot, the AI Insider for AI Trends and a regular contributor
I was driving my car the other day and came to an intersection that allowed me to make a left turn, but there wasn’t protective pocket and nor was there a green arrow available. Instead, this was a rather harrowing circumstance of occupying a lane that allowed cars to either go straight or make a left turn, and that permitted a left turn once oncoming traffic provided a break significant enough to make the left turn. A car ahead of me was also trying to make this death defying left turn. I could see her inch her car further forward into the intersection and was obviously hopeful that if she wasn’t able to make the left turn during the green light that she could at least do so once the intersection went to yellow. The oncoming traffic was streaming continuously and there wasn’t any break that she could try to dive into.
The light went to yellow and the oncoming traffic was determined to get as many cars straight through as possible. Finally, the light went entirely to red and there were still two oncoming cars that opted to scream through the intersection at rocket like speeds. The left turning car ahead of me was in for trouble since it couldn’t make the left turn yet, and now there was an onslaught of perpendicular traffic that was eager to use their green light to proceed. After some honking of horns and waving of fingers, the left turner made the turn and probably was still shaking from fright and of being touched by the icy grip of the grim reaper.
There are some that say we should ban all left turns. That’s right, ban them. According to the National Highway Traffic Safety Association (NHTSA), whenever there is a crash involving cross of paths, it is about half the time involving a left turn, while only about 6 percent of the time does it involve a right turn. New York City says that left-hand turns are about three times as likely to involve an injury or death for pedestrians as do right-hand turns. The NHTSA statistics indicate that over a third of motorcycle fatalities involve a car that was making a left-hand turn that occurred in front of an oncoming motorcycle. Overall, these rather depressing statistics do seem to suggest that left turns are the bane of our existence.
Designers of roads and traffic pathways are urged to avoid creating left turns and there are calls afoot by regulators to reduce the number of left turns in our transportation infrastructure. No left turns would mean that all of these incidents and accidents involving left turns would disappear. No need to fret in those left turn lanes anymore. No need to have a special green arrow to approve making a left turn at an intersection. In theory, we would only make right turns. That’s it, the right turns wins out. Go, right turns, go.
An often-cited aspect about the dreaded left turn is that many of the logistics companies abide by the rule-of-thumb to avoid left turns. For example, UPS is widely known for having instituted a practice of avoiding left turns. Keep in mind though that they did not ban entirely left turns. It is an urban myth that some believe that UPS executives have decreed that their drivers are to never make any left turns. This notion is entirely fiction. Instead, UPS uses their specialized GPS mapping system to minimize left turns when appropriate to do so.
What’s also interesting about this minimization of left turns is that it is not solely due to the accident rates associated with left turns. According to most studies of left turns, the time you spend waiting to make a left turn is relatively high, and so you are then delayed in getting to your destination and your vehicle uses more fuel. For transport and logistics companies like UPS, time is money, and so is fuel. If they can reduce the amount of time to deliver a package, and reduce fuel consumption, it will make them more profitable.
Why do we care here about left turns? When developing a self-driving car, one question that keeps being raised is whether or not a self-driving car should be making left turns. It might be safer for self-driving cars if they did not make left turns. Safer for the self-driving car and its occupants, and safer for other cars on the road, and safer for pedestrians too. Furthermore, using the UPS example, presumably the self-driving car might get faster to its destination and use less fuel, if it did not use left turns (though the driving practices of delivering packages is not the same as everyday driving patterns).
I’ve even heard some say that self-driving cars should be forced to never use left turns. A law should be passed that would ban any self-driving car from ever making a left turn, say some alarmists. These extremists even want that the AI of the self-driving should not know how to make a left turn, which then will ensure that the self-driving doesn’t inadvertently make a left turn, ever.
Crazy talk! That’s right, this is all crazy talk. Regulating self-driving cars to prevent them from making left turns, this is nutty. Purposely omitting the left turn capabilities from the AI of the self-driving car, even nuttier. Left turns are here and will continue to be here for many, many years to come. In fact, if you want to believe the utopian self-driving world of the future, which is when all cars on the road will only be self-driving cars, and these self-driving cars will communicate with each other via V2V (vehicle-to-vehicle communications), we can have left turns aplenty. That’s because in this utopian perfect world the self-driving cars will all politely talk with each other and coordinate their movement. Thus, a left turn will be a piece of cake. The ongoing traffic will talk with the self-driving car that is making the left turn, and the oncoming self-driving cars will ensure that a proper break in traffic allows the left turner to proceed without any worry.
But, let’s get back to reality. The AI of self-driving car needs to know how to make left turns. Left turns are an essential maneuver in driving a car. Now, this does not mean that the AI needs to always rely upon left turns. Similar to how UPS urges drivers to avoid left turns when feasible, the AI can be programmed to try and avoid left turns that are especially dicey. Based on history of the roads that the self-driving car is driving on, the AI can judge the risks associated with any given left turn and then opt to make the left turn or not. Human drivers also make this same calculation, and often choose to avoid a left turn or do something else about a particularly gnarly left turn.
Here’s a left turn example that I encountered just last week. I came up to an intersection that my GPS said I need to make a left turn as part of my path toward my destination. Cars were already quite backed-up and overflowing out of the left turn pocket. These cars protruded out into the normal lanes of traffic. Only about one or at most two cars at a time were able to make the left turn, during the intersection green-yellow-red light cycles. I realized that if I sat in this long line of cars that were awaiting making the left turn that it would take maybe an added ten minutes just to make that one turn.
I looked ahead and saw another left turn just a short distance away. There wasn’t any traffic sitting in that left turn lane. So, I drove up sneakily about a quarter block ahead and used that left turn to make my needed turn. Yes, it put me a little away from the street that I was supposed to be on, but with an additional turn I then got onto that street. I am sure that I easily beat the time that would have taken had I stayed in that earlier overly crowded left turn lane.
The point is that yes, I did make a left turn, but used one that was faster and safer to make. Judicious use of left turns is the mantra for any human driver, and likewise should be the mantra for any sophisticated AI that is doing the self-driving of a car. The existing situation surrounding any given left turn should be processed and a determination made about whether to make that left turn or not. Blindly making a left turn simply because the GPS says to make a left turn is not very smart. We want the AI to go beyond the norm of a typical GPS and add some smarts to making turns, especially for left turns that are in bad situations.
We have been analyzing left turns as part of our efforts at our Cybernetic Self-Driving Car Lab and devising self-driving car modules that specialize in dealing with left turns.
What’s a bad situation for a left turn?
Here’s some clues:
- No left turn pocket
- Left turn restricts the forward flow of traffic
- Left turn at a major intersection
- Lots of traffic involving the left turn
- Nighttime tends to be dicey versus daytime
- Visibility and weather conditions obscuring the roadway
- Left turns known for accidents
- Places where lots of pedestrians are nearby the left turn
- And so on
The roadways are designed to allow for Protected Permitted Left-Turn (PPLT) in certain areas. In theory, the local transportation authority is supposed to be monitoring these left turns and adjust the left turns or remove them if they are dangerously high in rate of accidents. Our system taps into these statistics to also try and gauge how lethal a particular left turn might be. There is also the LTAP/OD, which means to be aware of the Left Turn Access Path (LTAP) and the Opposite Direction (OD) of traffic.
For human drivers, they are taught to follow certain kinds of steps when making a left turn. Experienced drivers tend to do these steps and are not even aware that they are doing so. Though, some experienced drivers have over time opted to neglect the steps and so they are making left turns perhaps more on a hunch than by actually abiding by advised steps. If you watch a teenager learning to drive, you can see the steps since they are often doing them for the first time and make overt actions corresponding to those steps.
Here’s the typical steps involved:
- Have the AI move the self-driving car into the left turn lane or hug the center divider line if no left turn lane exists
- Have the self-driving car start signaling before it makes the left turn, which lets traffic around it know the intentions of the AI
- Reduce the self-driving car speed as it comes toward the left turn
- Come to a stop behind the limit line if it cannot in one swoop make the left turn
- Scan the oncoming traffic to identify a break in traffic using radar, cameras, LIDAR
- Gauge the time and distance needed to make the left turn
- Ensure that there will be sufficient clearance regarding oncoming traffic
- Scan for pedestrians and ensure that the turn won’t conflict with their movement
- Scan for bicyclists and ensure that the turn won’t conflict with their movement
- Scan for motorcyclists that are oncoming and that might conflict with the turn
- Determine the safety factor associated with when to make the turn
- Execute the turn once the safety factor is at an acceptable threshold
- Don’t cut the corner on the left turn
- Don’t go too soon
- Don’t go to late
- During the turn, continue to monitor all sensors and adjust as needed
- Keep the wheels straight prior to making the turn (in case hit from behind)
- Judge whether the car behind is going to try and make the turn too
These approximately twenty steps are just a tip of the iceberg of the AI cognitive effort involved in making a left turn. It needs to anticipate the left turn. It needs to plan for the left turn. It needs to be collecting sensory data and use it to assess the nature of the left turn. It needs to continually adjust to the evolving situation about the left turn. It needs to do re-planning as needed. It needs to execute the left turn. During the execution, it needs to continue collecting sensory data and be ready to adjust the driving action.
Situational awareness is crucial. Is this a left turn in a suburb or in the inner city? Is the road well paved and designated or is it a hidden path? Do the nearby drivers seem to be driving carefully or with abandon? Does the time of day make a difference? Does the day of the week make a difference? What is the volume of traffic? What are the speeds of the traffic? Is it at a controlled intersection? Is it a left turn in the middle of the road? Etc.
This is where collecting lots of self-driving car data is especially handy. If there are hundreds of other self-driving cars that have tried to make that same left turn, the system can use that data. By applying machine learning, it can ascertain the nuances associated with successfully making that left turn. The machine learning might discover that there is a pattern to always going quickly or instead always waiting until the last car traverses past.
There are other twists to the left turn cognitive effort too. Certain kinds of tactics and strategies apply to a left turn from a two-way street onto a two-way street. There are other kinds of tactics and strategies for making a left turn from a two-way street onto a one-way street. Likewise, for a one-way street onto a one-way street, and for a one-way street onto a two-way street.
Most self-driving cars are already able to do left turns in a rudimentary fashion. Tending to be extremely cautious, you can pretty much know when a self-driving car is making a left turn, since it is about as timid as a teenage driver learning to drive. As mentioned in my column about defensive driving and also the column about the foibles of human drivers, there are human drivers that will “game” a self-driving car and play chicken with it. If you were in an oncoming human driven car and saw that a self-driving car was readying to make a left turn, you could pretty much “know” that the self-driving car is going to wait for you to pass before it opts to make the left turn.
Those squeaky close left turns that human drivers make are not how the self-driving cars have been programmed. Instead, most of the self-driving cars are programmed to only make the left turn when it seems abundantly clear that the left turn can be made. Some believe that we will gradually see the self-driving cars get more akin to those squeaky close driving aspects by having used their machine learning to refine making left turns. In essence, if a given left turn has been undertaken thousands of times by the collective of other self-driving cars, then any particular self-driving car can use the pattern and optimize for traversing that left turn.
Let’s at least dispense with the idea that left turns should never be done by self-driving cars. This is going to continue to come up and I am expecting that “journalists” that don’t know much about self-driving cars are going to wring their hands and argue for doing away with left turns by self-driving cars. I would even predict that we’ll soon be having an accident involving a self-driving car while making a left turn, and this will get the mass media all charged up about left turns. Regulators that want to look good to their constituents might jump on the “no left turns for self-driving cars” bandwagon. I say, don’t toss out the baby with the bath water.
Making left turns is part of the real-world of driving. We need to ensure that self-driving cars can make left turns. We need to continuously improve how self-driving cars make left turns, and strive to aim for as much safety as we can. No matter how safe we aim, in the end, the self-driving cars are mixing with human driven cars and the combined unpredictable nature of the real-world is going to have self-driving cars get into accidents. The utopian world is not here yet. Meanwhile, let’s all drive safe out there.
This content is original to AI Trends.