Solving the Roundabout Traversal Problem for Self-Driving Cars


By Dr. Lance B. Eliot, the AI Trends Insider

Help, I am stuck in a roundabout and can’t get out. For those of you that have ever driven into a roundabout, often also known as a traffic circle, a road circle, and sometimes a rotary, they can be devilish to navigate. Typically, a novice driver finds them to be frightening and a real-world version of a crazy bumper car mad dash. Seasoned drivers like to think that they have mastered the roundabout and so act like it is a breeze to traverse one. Even the seasoned drivers though are at times thrown for a loop and find themselves baffled and tortured by the roundabout. If you get enough drivers going through a roundabout and if they are all behaving badly, you find yourself wishing you had gone some other path and had avoided the dreaded roundabout.

As the head of the Cybernetics Self-Driving Car Institute, we’ve been developing techniques and software to enable self-driving cars to properly traverse a roundabout.

Most of today’s self-driving cars hand the driving back to the human driver when encountering a roundabout.  That’s if the self-driving car even realizes that a roundabout is about to occur or occurring. Some self-driving cars head into a roundabout without the realization that it is a roundabout. The AI of the self-driving car is often ill-prepared for the specific dynamics of a roundabout. As such, the AI either struggles to make it through the traffic circle safely, or at the last moment is gives up and tosses control of the vehicle to the human driver.  This is not only dangerous for the human driver and passengers, since the act of taking over control suddenly can be disruptive and create confusion for the human driver, it also violates the principle of the Level 5 true self-driving car which is that the AI must be able to do whatever driving a human driver could do.  In that sense, we need to have a solution for roundabouts if we are going to achieve Level 5 self-driving cars, which we are still some distance from achieving (see my column on the Richter scale for self-driving cars).

For those of you that have rarely if ever encountered a roundabout, they aren’t especially common in the United States, numbering an estimated 3,000 or so across the entire country. In contrast, if you’ve ever been to Europe, and especially France, you’d have seen quite a number of these roundabouts. France has the fame of having the most roundabouts in the world, numbering around 30,000, so about ten times the number of roundabouts in the entire United States. Given the size of France in comparison to the United States, the per capita or per square mile ratio of the number of roundabouts is extremely high in France.

The United States has had roundabouts since before the advent of the car.  Some consider the most famous and early notable roundabout to be the Columbus Circle in New York City (shout out to NYC!), which was designed and put into place around 1905. There are numerous studies about roundabouts and all sorts of engineering aspects underlying them.  Traffic engineers tend to like the roundabout. It is a means to regulate traffic flow. In theory, it is as safe if not safer than a conventional intersection, it flows traffic more smoothly, it reduces wasted fuel consumption in comparison to having cars sit at a traffic light of a conventional intersection, it produces less pollutants because again the cars aren’t sitting at an intersection, and otherwise it is just kind of cool.  The Traffic Research Laboratory (TRL) provides very helpful analyses of roundabouts, particularly the ones in the United States, and provides a somewhat definitive guide to the topic.

There are a myriad of variants of a roundabout. There are mini-roundabouts, magical roundabouts, there are roundabouts in city settings, ones in suburban settings, ones nearby to schools, etc. I won’t cover all the variants here, and focus instead on the overall model of a roundabout.  If a self-driving car can handle a modestly complex generic roundabout, it can then tailor to whatever specific circumstance it faces for a particular roundabout.  But, if the AI lacks any knowledge of a roundabout, its overall driving rules and capabilities are likely going to get it into trouble and so the need for a specialized component to guide it during the roundabout traversal.

You might wonder why traffic engineers believe that a roundabout is safer than conventional intersections. When you have wandered into a roundabout with lots of other traffic, it often seems to be like a swirling pit of man-eating sharks, and you might think it has to be the least safe way to ever design any kind of traffic flow. According to the stats collected by the TRL, they claim that it is actually much safer than conventional intersections. Part of the logic is that when you come to a conventional intersection, you have the chance of ramming other cars at right-angles, you have a chance at hitting other cars at a heads-on position, and you have a chance at hitting other cars during a left turn. All of those aspects are generally eliminated via the roundabout.  For the roundabout, you mainly have the chances of doing glancing blows off of other vehicles. Rarely do those glancing blows then lead into a full-fledged derby style cascade of cars knocking into each other.  Of course, you also have the risk of rear ending other cars, such as when you first try to enter into the traffic circle, and when you try to exit the traffic circle.  These roundabouts are certainly not risk free, even if in theory “safer” than conventional intersections.

What constitutes a roundabout?  The usual roundabout is composed of an inner core that is an island, upon which no cars are allowed to go. Traffic flowing to the roundabout usually comes from 360 degrees around the island.  The lanes flowing into the circular area are at staggered positions around the circle. Roundabouts typically will have two to four such entrances into the circular flow.  There are usually concentric circular lanes surrounding the roundabout.  Sometimes there is only one such lane, into which the entering traffic must squeeze. More often, there are two or more circular lanes, referred to as multi-lane roundabouts. Traffic flowing around the circle jockeys into and out of these concentric circular lanes.  There are inner lanes that are closest to the island, and outer lanes.

You have to somehow enter into the roundabout, and somehow exit from the roundabout. The entrance lanes are usually posted with a Yield sign and the driver must yield to the traffic already flowing around the circle. In addition to entrance lanes, there are exit lanes. The exit lanes will sometimes be separated from the entrance lanes and are their own distinct lanes.  More often, the entrance and exit lanes are aligned with each other, but usually there is a separator to try and ensure that exiting traffic doesn’t inadvertently try to go into an entrance lane. Between the entrance lane and an exit lane there are often splitters that keep those separated from each other.  A splitter is an area marked for cars not to go onto it, which can be an actual raised median or it can be just painted onto the ground.

The core island of the roundabout is sometimes just a painted marked area, but more often it is a raised island and has something on it.  It might be covered with grass or other low-to-ground aspects, allowing the drivers to see across the island and be able to detect traffic that is on the other side of the roundabout.  Most cities though try to make the island into something noticeable and attractive, and so they put artwork there, such as statue of the founding figures of the city, or they put a fountain that is spitting out water and provides a nice respite from the visual dullness of cars flowing around and around of the roundabout.

For self-driving cars, this aspect that the island has something on it can be an added difficulty factor when trying to traverse the roundabout.  The sensors of the self-driving car are not readily able to penetrate whatever is on the island, and so it limits an ability to predict traffic patterns.  The cameras of the self-driving car and the LIDAR and radar are only able to get a partial indication of what traffic is on the other side of the island, and not able to gain a full sense of the other cars that are then coming ultimately toward the self-driving car via the concentric circles.  It is still worthwhile to try and get that data in real-time and analyze it, but the AI of the self-driving car has to assume that the data will be noisy, obscured, and only provide at best partial information about what is taking place on the other side of the island.

Let’s now take a look at how the self-driving car can properly traverse a roundabout.

First, the AI should be doing a look-ahead to anticipate a roundabout. Via GPS, the AI can potentially already be aware that a roundabout is part of the route being traveled.  If so, the AI might determine that trying to deal with the roundabout is not worth the trouble and instead opt to find an alternative route via the GPS that eliminates the need to traverse the roundabout.  Assuming that the alternatives do not otherwise significantly raise the time required for travel or have other adverse consequences, it is often feasible to simply skirt around a roundabout and take some other viable path.

Suppose there is no other viable path, or suppose that the GPS is either not available or has failed to realize that a roundabout is soon to be encountered, the AI needs to be ready to handle the roundabout. Often, there are street signs that indicate a roundabout is upcoming (see my column on self-driving cars and reading street signs). The AI might be forewarned that a roundabout is imminent as a result of reading a street sign, and so either opt to avoid the roundabout at the last minute and find an alternative path, or at least begin to prepare for the advent of the roundabout.

Upon arriving at an entrance to the roundabout, the AI needs to use its sensors to ascertain how many lanes are combined into this particular entrance. If there is only one lane, then the self-driving car has a simplified task since it only needs to focus on getting into the circular traffic that is flowing around the roundabout.  Things become more complex when there are two or more lanes that combine into the particular entrance that the self-driving car is using.  In such a case, the AI needs to now observe the cars that are in the aligned lanes of the entrance. These cars can be a mixture of human driven cars and self-driving cars, either of which will be potentially hard to predict in terms of driving behavior.

The next aspects of traversal can be potentially guided by machine learning, if the self-driving car has prior data associated with this roundabout, and if the prior data has been analyzed by the machine learning to identify patterns for best navigating the roundabout. This could be the case if there are many self-driving cars that have been using the roundabout and then feeding their data into a global data collection for use by the machine learning.  For now, we’ll assume this is unlikely to be available and that the AI will need to directly try to mentally muscle its way through the roundabout.

When two or more lanes exist in one entrance, the cars at the front of the line that are seeking to enter into the circular flow of traffic will tend to jockey to see which makes the first move into traffic. Furthermore, if say another car is immediately to the left of the self-driving car, and if that car decides to dart into traffic, it potentially cuts off the self-driving car from being able to make its move into traffic. The AI needs to detect the cars aligned in the entrance and ascertain the nature of their behavior as to whether aggressively inching forward or being at a halt.

This brings up another aspect namely that if the conditions are proper, the self-driving car can directly feed into the circular flow of traffic, without having to come to halt or even potentially slow down. As the self-driving car makes its way into the entrance, the normal driving control would be to slow down so as to allow for judging of making a move into the circular flow of traffic. In fact, roundabouts are known for their “traffic calming” effect in that the cars coming to the entrances will usually decrease speeds in order to obey a yield to the flowing traffic.

There are though sometimes “perfect conditions” in which the self-driving car can smoothly come to the entrance and then continue unabated into the circular flow.  For example, if there is no other traffic already on the circular flow, or if the traffic in the circular flow has provided an appropriate gap, the self-driving car can make one continues shift from the entrance and into the outer lane of the circular flow.

One of the most arduous analyses for the roundabout indeed involves finding a gap in the circular flow that will safely allow the self-driving car to makes its way into the circular flow. Humans often have a tendency to make a gap, called gap-acceptance, allowing an entering car to gain entry into the circular flow.  Some drivers while driving the circular flow will purposely slow down or speed-up in anticipation of allowing a car to move from the entrance into the flow.  Of course, there are other drivers that either don’t care about providing a gap, or even take the opposite stance and intentionally cut-off the entrant from entering into the flow.  This varies depending upon the time of day, the day of the week, the volume and velocity of traffic, and the local cultural norms concerning the particular roundabout.

Once into the circular flow, the self-driving car usually starts at the outer lane of the concentric circles, assuming there is more than one such lane. Part of the analysis now becomes whether the AI should guide the self-driving car toward the inner lane of the circular flow or remain in an outer lane. If the exit for the self-driving car is quickly approaching, the AI should keep the self-driving car in the outer lane. If the exit is on the other side of the circle, making movement into the inner lane can be beneficial, and avoid the rather chaotic aspects that occur on the outer lane. The outer lane is a continual frenzy of cars gaining entrance and exiting, and there is at times a greater calm to be found by moving into the inner lane.

As the self-driving car nears the desired exit, it needs to make its way to the outer lane, assuming it is not already there. If other cars are unwilling to allow the self-driving car to jockey out to the outer lane, it is conceivable that one full iteration of the circle might be needed to accommodate getting over to the needed outer lane. Observations of human drivers shows that they will often take this same tact. Namely, they have become trapped in the inner lane and the surrounding traffic won’t let them safely make their way to the outer lane.  This can be the fault of the human driver for not having started toward the outer lane soon enough to make it to the desired exit.  The self-driving car should be able to appropriately ascertain the distances and the lanes to early enough start to make a move toward the outer lane and ultimately be in a position to make the exit.

But, this transition from the inner lane to the outer lane is contingent on the behavior of the other drivers. If the other drivers are stubbornly not allowing the inner lane self-driving car to make its way to the outer lane, it might be necessary to make a full revolution and therefore seek a more favorable condition on the next round. Presumably, the cars that were purposely blocking the move will by now have exited themselves from the circular flow of traffic, and hopefully other remaining cars will allow for a movement out of the inner lane.

The behavioral aspects of drivers in roundabouts is a crucial element in traversing the roundabout. Legally, most roundabouts are designated such that the cars entering into the circle must yield to those cars already flowing in the circle. What happens in actual practice can differ significantly.  In fact, “priority reversal” occurs quite often, whereby the cars entering into the circle don’t wait their turn, and instead they force themselves into the circle.  This causes cars that are already in the circular flow to have to yield to those brash entrants.  This can become a dangerous game, since the entrants are nearly daring the already flowing cars to hit them, and in this game of chicken somebody has to give way.

In terms of the exit from the roundabout, the self-driving car needs to make its way to the outer lane and then flow into the desired exit. This can be made more arduous by other cars that are trying to get past the exiting car and at times block the exit for the exiting car.  Once again, if needed, the self-driving car might need to abandon its attempt to achieve the exit and make another loop around the roundabout. The nature of the allowed aggressiveness of the self-driving car will partially determine whether it can quickly make the loop and exit, or whether it will need to possibly make multiple loops.

Most of the self-driving cars today are programmed to be the “old granny” style driver. The AI takes an ultra-conservative approach to driving the car. Furthermore, any more aggressive actions are simply handed over to the human driver, rather than having the self-driving car take such actions. As I’ve mentioned in prior columns, this timid style is sensible during the existing era of trying to get self-driving cars to be simplistic drivers on the roads, but eventually we’ll need to ratchet up the aggressiveness of self-driving cars.

After the self-driving car has made its way to the desired exit, there is usually a need to speed-up at that point.  The exits from most roundabouts tend to flow back into normal street traffic, and so the self-driving car needs to accelerate from what was likely a slower speed during the roundabout and into a faster speed to match with the traffic flow beyond the exit. It is also conceivable that the self-driving car might have gotten itself mired in the roundabout and taken an exit by mistake, which, if so, the self-driving car would now need to recalculate how to get back onto whatever path was supposed to be taken. We are all used to this when we hear our GPS say “recalculating” after we’ve taken the wrong turn. I know that it seems unimaginable to think that the self-driving car would have taken the wrong exit, since we expect that a self-driving car is to be perfect, but, if the roundabout has become high risk it might make better sense to get out of it, rather than pursuing “perfection” for the desired exit and yet put the occupants at undue risk to do so.

In spite of the overall notion that roundabouts are safer than conventional intersections, there are certainly many ways in which a self-driving car can get itself into hot water in a roundabout. Upon entering into the roundabout, the self-driving car could inadvertently rear-end another car that has tried to cut-off the self-driving car or that was in the loop and suddenly decide to brake. Once inside the flow of the roundabout, other cars trying to weave into and out of the concentric lanes can readily strike the self-driving car.  Cars coming into the roundabout that are behind a self-driving car might become impatient and rear-end the self-driving car.  And, as mentioned earlier, there are lots of chances for having glancing blows among cars that are in the concentric circles.

Besides developing AI software for dealing with the particulars of roundabouts, we are also doing simulations to gauge how well the AI can deal with the uncertainties associated with roundabouts. We purposely feed “human driven” cars into the roundabout that consist of humans that are the Indy race car driver type that thinks the roundabout is fun and intended for battle, and we feed into it the timid drivers that drive into a roundabout once a year or once in a lifetime and mess-up the flow by going slowly and not abiding by the flow of traffic. These real-world simulations help the AI to learn how to deal with the variety of human drivers that will be encountered while traversing a roundabout. Drive safely out there.

This content is original to AI Trends.