0:00
/
0:00
AI Self-Driving Cars Coping With Sudden Lane Changes
AI Self-Driving Cars Coping With Sudden Lane Changes
0:00
0:00
0:00
0:00
About this episode
Exploring the complexities of lane changes for self-driving cars, Dr. Lancelot delves into the psychological and mathematical challenges involved. He discusses how human drivers navigate lane changes, weighing factors like speed, time, and the behavior of other drivers. The episode highlights the limitations of current AI in making lane changes, often resulting in overly cautious maneuvers. With insights into buffer zones and real-time calculations, the discussion emphasizes the ongoing development needed for AI to match human driving instincts in dynamic traffic situations.
Topics:
lane changes
ai behavior
human psychology
speed factors
buffer zones
real-time calculations
traffic dynamics
safety considerations
Dr. Eliot explains how AI self-driving cars cope with sudden lane changes. See his Forbes column for further info: https://www.forbes.com/sites/lanceeliot/
Select text to request an explanation
Hi, I'm Dr. Lantz, and welcome to my podcast series about self-driving cars.
In this episode, I'll be discussing the topic of changing lanes and self-driving cars.
If you become interested in learning more about self-driving cars, please see my website
www.ai-self-driving-cars.guru for further information.
Okay, let's get started.
Hey buddy, pick a lane and stick with it.
You've likely thought about or spoken those remarks about roadway lanes.
We live in a world of lanes.
Until the day arrives that we have flying cars, we all need to be cognizant of roadway
lanes.
It's kind of amazing when you think about it that people are willing to most of
the time obey the various lanes.
Thanks goes to all you drivers for a general consensus on sanity when driving.
For teenage drivers that are learning to drive, they often find that changing lanes is one
of the scariest aspects of driving a car.
Should I go now or should I wait?
Can I make the lane change or will I cut off another car?
Is that car going to allow me into that lane?
Should I speed up?
Should I slow down?
I remember when I was first learning to drive a car, we had five high school
students in my driver ed class and for which scarily we always all drove in one car with
the driving instructor.
Each student would get about ten minutes of driving time per instructional trip.
Changing lanes was a dangerous act.
Think about what can go wrong.
You can start the lane change and perhaps not realize there's a car directly aside
of you and then hit that car.
Or maybe that car realizes you are mistakenly getting into their lane and so they hit
you.
Or they veer and hit another car as caused by your having started your lane change.
Let's assume you're good enough that you don't start the lane change when there's
a car next to you.
But once you engage in the move, a car suddenly appears in that lane.
Once again the chances of getting hit or hitting the other car or starting a cascading accident
that will possibly injure or even kill other people is all heightened.
There's some key factors involved in the lane change act.
One of those factors is speed.
lane changes at low speed is more likely to have less adverse consequences.
If you hit another car and you're both going say five miles per hour, certainly it's not
a good thing.
But hitting another car while going 70 miles per hour certainly apt to have more dire consequences.
The faster speeds tend to mean less reaction time and so less chance of avoiding a collision.
The faster speeds tend to mean greater damage to the cars and greater chances of injury
to the human occupants.
Faster speeds tend to mean that the instant will turn into a domino and lead to a
multitude of crashes.
Believe it or not I know one driver that genuinely believes that higher speeds are safer for making
lane changes.
His logic is the length of time involved in actually undertaking the lane change is reduced.
At 70 miles per hour he says it's a blink of the eye when you make a lane change.
It's like being the flash and it just happens so quickly that essentially nothing
can go wrong.
At slower speeds he believes it is worse since you take too long to make a lane
change.
This longer time frame increases the opportunity for a car collusion in his mind.
All he can say is wow what a way of thinking.
I'm glad that he and I are rarely on the roadways at the same time.
This also brings up the useful facet that we have both your speed as a factor and also
time as a factor.
You need to know what your speed is assuming you want to make a reasonably well executed
lane change.
You need to know how much time it will take to make your lane change.
You need to estimate how much time there will be for a gap to exist in the other
lane such that you can get into that gap.
You need to calculate whether you need to speed up or slow down to make it into
that gap.
The gap is both a physical gap in terms of physical space where your car needs to
fit and that no other car occupies and it's a time gap in that it needs to happen
at a point in time during which there isn't another car there.
You might be tempted to think that changing lanes is merely a series of algorithmic
like calculations that we humans are making.
I would suggest that's oversimplifying the beauty, art and craft of lane
changing.
You need to be part psychologist too.
Will that other driver in the lane next to me be willing to let me in?
There's a momentary gap in space and time right now but the next few seconds it could
disappear.
The other car might suddenly accelerate and end up in the spot that I think I want
to occupy.
Your lane change is actually a forecast about the future.
You are trying to predict what the road to a situation will be in a few
seconds from now.
This involves making educated guesses.
You can't usually know for a certainty what the future is going to hold.
Now yes, there are situations involving easy lane change actions.
If I'm on the open highway and there's not a car around me for miles, I can change lanes
to my heart's content.
For ease of discussing lane changes, let's refer to the driver that is initiating the
lane change as the activator.
The car that is in the other lane that might be impacted by the activator will be referred
to as the responder.
I'll take the core elements first, namely, assume that we have one activator and
one responder.
That being said, this can readily become much more complex by adding multiple responders or
multiple activators.
The most reduced and simplest instance would be one activator and zero responders, which
is akin to my example that you can just change lanes readily when there's no other
cars anywhere near you.
Consider first the speeds involved.
In the first case, imagine the case of the activator car and the responder car both
going at high speeds, and we'll assume for a moment for simplicity that they're going
at roughly the same speed.
Generally the time to make the lane change will be compressed because of the high speeds.
Generally the risk factor will be higher in spite of what my friend believes, since
they're both going at high speeds.
In the second case, both the activator car and the responder car let's pretend are
going at low speeds.
We'll assume for simplicity again roughly at the same speed.
During the time to make the lane change will certainly be longer than the higher
speeds instance, and the risk factor tends to be lower.
In the other two cases there's a disparity between the speeds of the two cars.
This generally creates a situation, less clear cut, when both cars are not at the same speeds.
In the instance of the activator going at let's say a high speed and merging into
a lane with a responder going at a lower speed, presumably the activator has some
advantage since they can rarely get ahead of the responder and potentially avoid having
the responder strike them from the rear.
In comparison, when the activator is going at a lower speed and tries to merge in front
of a car that's going at a higher speed, this can be a quite heightened risk since
the responder car is going to have to likely react.
This brings up another factor, namely the ability to accelerate and decelerate or
brake.
We all know that when making a lane change you don't necessarily maintain your existing
speed.
You can but often you speed up to make the lane change or you alternately slow
down.
Likewise, the responder might have to react by speeding up or slowing down.
Let's return to the psychology of the lane change act.
The activator does not particularly know how the responder will react to the lane
change.
It could be that the responder might suddenly speed up and thus the prediction
about the future of physical time gap is no longer what was predicted.
Sometimes the responder doesn't want to let the other car into their lane.
Sometimes the responder is oblivious to the lane change and just decided to speed
up due to a completely unrelated aspect.
How does another driver know that you're trying to undertake a lane change?
Well, legally you're supposed to signal that you want to make a lane change.
Much of the time people do indeed turn on their signal blinker.
This though can be slippery.
Some lane changers will turn on their signal after they've already intruded into the
other lane.
This could be because they forgot to use the signal and suddenly remembered or it
could be that they didn't want to give a heads up beforehand and then as a kind
of pretend afterwards turn on their signal to act as though they had properly
signaled to make a lane change.
In that sense lane chains also involve head fakes and head fake in this context is
when you begin to move your car towards a lane change and though you haven't made
a full commitment to it you are signaling that you're making a lane change.
This can occur either while using your turn signal or not even using the turn
signal.
So we have the potential of a head fake to showcase an upcoming lane change and
we also have the use of an actual turn signal.
What does all this have to do with AI, artificial intelligence and self-driving
cars?
The Cybernetic AI Self-Driving Car Institute were developing AI software for
self-driving cars and one of the areas that's rapidly being advanced involves
the ability for a self-driving car to make lane changes.
For most of the self-driving cars of today if you ever watch them make a
lane change you might almost laugh at what you see.
Generally they make a lane change like a timid teenage novice driver would do.
The AI of today will only make the lane change if it seems absolutely abundantly
the case that the lane change can be made with a great deal of safety.
As such the AI often starts doing a lane change a lot sooner than most human
drivers would.
Also the AI self-driving car tends to often disrupt traffic flow when making
lane change which is not by design but by the aspect that imagine if you had
a teenage driver making a lane change on a busy freeway.
The other drivers will all pretty quickly size up that driver and either
give them a much wider berth or try to take advantage of them.
This is kind of like playing a game of poker and a rubbe sits down to
play.
Everyone else knows how to play poker will try to every trick in the
book on that rubbe.
I realize that some of you are going to say well there's no need for an AI
self-driving car to worry about driving involving humans because we're going
to have all self-driving cars on the road someday.
Wake up that's going to be a long long time before that happens.
There are about 200 million conventional cars in the United States alone
we're going to have self-driving cars mixing with human driven cars for a
very long time.
Recall that I mentioned that making lane changes is not just the act of
making calculations.
It also involves the psychology of the other drivers.
For human drivers right now they can readily guess the psychology of the
AI self-driving cars.
Just assume that the AI will pretty much do whatever is the most
conservative and novice act of changing lanes and that's it.
Having to figure out other humans is much more complex though
seasoned drivers have already assessed how to do so generally.
Season drivers look at not only the behavior exhibited by the other
car and the other car driver but also often include looking
at other aspects such as what does the driver look like?
What does their car look like?
Is their car in good shape or bad shape and so on.
The AI of today's AI self-driving cars doesn't consider any of those
other aspects.
Instead the AI simplistically detects that there's another object that's
moving at such and such speed and otherwise doesn't really care as to
the nature of the car nor the nature of the driver of that car or any of
the such fashions.
It is basically calculations-based.
Now some interesting new approaches to the mathematics of this include
incorporating what are called buffer zones.
Imagine that the AI self-driving car has kind of a cone around it which is a
buffer zone.
Other cars have likewise buffer zones around them.
This virtual buffer zone is kind of a pretend and that we're pretending
that the car is actually larger in a sense that it is.
You might think of this as a geofence placed around a car.
The buffer zone can be very wide like say we decide there's a pretend
area around a car that's four feet all around the car
or the buffer zone might be very tight such as we might pretend let's
say a foot or less in size.
When trying to make a lane change we'll mathematically consider that the
objects the cars are the sizes of their buffer zones.
One key principle is that we don't want to have buffer zones come in
contact with each other.
If the buffer zones come in contact it presumably implies a collision is
going to occur.
Now the question rises as to how risky we're willing to go.
If the buffer zone is four feet I presumably have some allowable
slippage that I can make the lane change and get say within three feet of that
other car and not actually hit it even though I have
punched into their buffer zone.
Mathematically we'd like to have the AI be able to prove
that collision avoidance will be preserved.
So we take the buffer zone of the activator and the buffer zone of the
responder we make assumptions about their speeds and
likely actions in the near future when the lane change will happen
and run through various calculations to see whether the lane change can be
made with a guarantee of no collision.
This is easy to do in the lab or via simulator.
In the real world this needs to be done in real time.
The question arises as to whether or not the calculations can be done quickly
enough to then make the decision for the lane change
and then do the lane change.
If the calculations take say three seconds too long
the opportunity for the lane change evaporates.
Furthermore the whole set of calculations is now worthless
since the circumstances have changed and a new set needs to be executed.
As much as possible it is best to pre-compute this.
You can have onboard the AI system lots of pre-computations done for varying
circumstances and thus instead of having to calculate at the moment
you can look up in various tables to see what those tables indicate.
This would be similar to chess playing.
You can either have a chess playing program that has to calculate all the
various permutations and options of each play
or you can have pre-stored templates that once the chess board is in a
certain configuration you can just look it up and know
what the next move would be.
We can do somewhat of the same with the AI self-driving car and lane changes.
Now not entirely though and so we are more likely to have at times
have to make some kind of raw calculations and if so
they need to be done in real time to match the time constraints of the
situation.
One recent approach by researchers at MIT involves calculating
a logistics function involving the buffer zones and the direction of speed
and so on and then combining that with a Laplace-Gas distribution
the so-called bell curve to do an on-the-fly estimation of the chances of
making the lane change and doing so with collision avoidance.
Efforts to solve this problem need to contend with the severe time constraints
and doing the calculations in real time along with considering how much
information is available and how reliable is that information.
The sensors of the self-driving car are providing data about what is around
the self-driving car.
You cannot assume that this is perfect information.
The sensors might be getting a lot of noise such as the cameras have
blurry images maybe due to weather conditions or maybe the radar is not
reflecting well off of other cars and so on.
The sensor fusion must be contending with conflicting and missing
information about the surroundings.
Changing lanes appears on the surface to be straightforward.
As a human if you are the activator you just glance over your
shoulder survey the scene maybe turn on your turn signal you
steer the car in the other lane and voila you're done.
The responder likewise simply has to notice that your car is providing
some indication that you're wanting to make a lane change
perhaps via your turn signal your head fake movement and then let you in.
I hope you realize now though that making lane changes a lot harder than that.
As usual it's one of those human learned aspects that after a while seems
effortless.
To get an AI system to do this with a car and in motion
and with all the variables in terms of surrounding traffic and the
roadway in the lanes it's a tough thing to do.
For the moment we've gotten the AI to make baby step lane changes.
By continuing to push forward in advance the techniques and software
we all aim to make lane changes as effortless as humans do.
This includes that at times the lane changes might be civil in nature
and in other cases more aggressive.
Soon enough you might find yourself saying hey buddy you cut me off
and then realize you should have said hey AI you cut me off.
Well thanks for listening again I'm Dr. Lancelot.
I hope that you found today's episode informative.
If you're interested in learning more about self-driving cars
please see my website www.ai-self-driving-cars.google.com
Request an explanation for:
Request an Explanation
Heard something you'd like explained? We'll add it to this episode.
Sign in to request explanations for terms you heard.
Want to learn more?
Browse our glossary for plain-English explanations of automotive terms, jargon, and concepts.