Brad Templeton Home
My Forbes Articles
(Book from 2008)
School of Fish Test
I'm proposing a sort of "Turing test" for autonomous vehicles which would serve to convince skeptics that they are safe.
In this test, a swarm of robocars of different types moves on a track. The skeptic is given a regular vehicle to drive in the swarm. If the vehicles pass the test, the skeptic can't hit a robocar no matter what they do. They can swerve, brake suddenly, even go against traffic and, so long as the traffic is not packed at an unsafe density, they can't hit a robocar or get one to hit them.
In an alternate version of the test, a pedestrian goes out into the flow and tries to touch a robocar. If they pass, no touch can be made. This is like a diver swimming with a school of fish. It's next to impossible to touch one of the fish. (Obviously until they can readily pass you won't get many volunteers.)
How to do it?
Here's my first blush at how you might build vehicles to pass this test. Each vehicle would operate on its own, as fish do, but they could communicate with other vehicles for extra information, and also sometimes trust other vehicles.
Everything on the road must be identified. Things would be classified into various groups:
For every other thing on the road, the robocar must forecast where it is going, and where it could go. Any untrusted vehicle will normally move along its current vector, but it might swerve or stop suddenly. Pedestrians will probably stay on sidewalks but they could step into the road. Animals and children might be more erratic. Rarely, a person might even fall down and be lying on the road. The more you know about something, the better you can predict where it can and can't be. However, bodies must follow the laws of physics. Even an unknown car can't stop or turn on a dime. A human can only run so fast or change direction so quickly.
The next question is to examine the zone into which the unknown body might move and consider whether it intersects where you plan to go. If it might, an escape route would be plotted. This escape route would depend on the superior reaction times of the computerized control system, and possibly superior maneuvering abilities. Road conditions must be factored in.
This requires there be spaces in the flow of traffic. If there is a space next to you, for example, you know you can turn into it if something unexpected happens in front of you. Doing so might remove a space somebody else depends on, but from a computer's standpoint, that all happens very slowly, and your erratic swerve to avoid an obstacle will cause other vehicles near you to likewise swerve and adjust the spaces.
If the vehicle in front of you decides to suddenly stop, you probably don't need to swerve. Your fast reaction time means you can start braking right away, and only a small gap is needed to avoid a collision. You will take into account whether the vehicle in front of you can brake more quickly than you, and also whether the vehicles behind you may not be able to brake as quickly, and adjust the gap at all times to allow you the room you need -- or depend on spaces to the left and right.
If your neighbours are trusted robocars, you can rely on them to move if you ask them to (or simply move towards them) in an emergency. They can report to you their own ability to swerve or stop, and this will allow you to plot emergency courses which involve swerving into where they currently are, but won't be by the time you get there. The fish of course all trust their fellow fish to react this way -- for them it is instinct.
The fish have an advantage our robocars don't -- they can expand to left and right in the open sea. So robocars will have to use traffic patterns that have even more gaps than fish need. That's fine -- in fact I believe the capacity of a road should really be defined not as how many cars can squeeze into it, but how many cars can fit while leaving enough gaps to always avoid problems. That's not so many gaps, though of course each emergency use of gap slightly increases the risk if another gap is needed right away.
If your neighbours are untrusted vehicles, you must leave wider berths. In some cases, you might want to accept that you'll swerve into a gap and the resulting closed gap will be too close for a short time. After you swerve into it, other vehicles will move appropriately to widen the spaces. This strategy poses a risk but allows more dense packing of vehicles.
Largely, there will be a calculation balancing safety with how densely traffic can be packed. Normally vehicles will maintain large gaps, far bigger than needed. They would only pack closer to tolerances during rush hour. We may decide that in order to get more rush hour capacity, we will shrink the safety tolerance that would protect a pedestrian who suddenly steps into the road. We certainly do that now -- any pedestrian who leaps into heavy traffic will get the blame for their demise.
At rush hour, cars would not be parked on busy streets, and thus margins could be left between sidewalks and normal traffic flow to allow more time to react to pedestrians. But again, to the computers, pedestrians will be very slow moving objects, easy to avoid.
In a pinch, robocars can even stop and swerve more quickly than human driven cars can. This of course could hurt their human passengers. Robocars however will know in advance of such moves, and could in making emergency moves deploy airbags or other safety technology at just the right time to avoid serious problems. This is not something you would want to happen regularly.
This system breaks down a bit in the event of dense, two-way traffic. In this case, it may be hard to leave enough space so that you can deal with an oncoming car suddenly swerving into your lanes. This could be prepared for in light traffic, but in heavy traffic it's challenging. In current society, we just accept this fact. We train our brains to ignore the otherwise frightening vehicles coming at us at high speed.
This can be mitigated with reasonable gaps between the two lanes, such as the left turn lanes we often use today. Robocars won't need left turn lanes with their ability to flow around stopped cars.
There may be some forms of extreme behaviour that may be impossible to avoid. Because of the need to pass and move at different speeds, there will sometimes be cars right next to you. If a vehicle right next to you decides to turn sharply in a way that would send it flying off the road, it may be very difficult to avoid. Normally this would require a crazy driver, but it could also happen due to a mechanical or control failure on another vehicle.
If gaps can not be left to prevent this, another alternative may be to do a controlled collision. In this case, you may decide to hit a vehicle that is careening out of control on the side before it can change course enough to get in front of you. This could be done by braking, accelerating and steering into the vehicle rather than avoiding it.
This might not actually be too dangerous if vehicles have airbags on the outside or a quickly inflatable rubber bumper on all sides of the vehicle. With a good enough rubber bumper a well predicted sideswipe might leave little or no damage and prevent the out of control car from going further out of control, possibly saving its occupants.
It's not out of the question that cooperating cars from all around an out of control car could move forward to contain it and guide it -- with some gentle impacts -- to safety. Even a car whose engine is at full could be stopped with cars boxing it left and right and multiple cars gently bumping into it from the front and applying brakes.
Airbags on the outside could also help in collision with pedestrians in extreme situations (such as those trying to commit suicide who jump from a hidden spot into traffic.) Unlike modern airbags which are triggered by impacts, a robocar will know well in advance (from a computer's reaction time standpoint) of any inevitable impact and where it will take place. It can fire an airbag at the ideal time to reduce injury. One can even imagine airbags which are shot out of a tube to expand some distance from the car if need be.
It is likely that once robocars can pass this test, pedestrians will come to expect it. Thus, we'll see people feeling they can jaywalk at will. Parents may be less wary and children may wander into the street more often. Children may even come to expect this, and play street hockey on live streets knowing the cars will zoom around them.
This does make the problem harder, and could also result in a lot of jostling for passengers in robocars, so it's an issue.
In California, technically if a pedestrian steps into the street, whether at a crosswalk or not, cars are required to stop. The pedestrian is jaywalking and would, in theory, get a ticket for this. In practice, cars don't stop, and pedestrians don't expect them to, and in practice, tickets are rarely written. People just work it out. However, robots won't be that flexible. It's hard to say what will happen. It could be that pedestrians who just walk out into the street at random will get photographed by the angry passengers in the swerving cars, and get tickets.
Another alternative would be dynamic crosswalks. If a pedestrian is able to convey they desire to cross the street at any random location, perhaps by pushing a button on their mobile phone, robocars heading to that location could be informed, and automatically adjust their pace and gaps to clear a dynamic space for the pedestrian to cross. With just a few seconds of warning, they could leave the right gap with no discomfort to their passengers.
Of course, it might take some time before people are willing to cross a street with a "virtual crosswalk" which is only void of cars in the section where they are walking. At first, we would make more traditional longer gaps, and as a result, require pedestrians to wait a bit longer for the clear signal. Though most people would not mind cars zooming through behind them, they would be scared of cars going through where they plan to be in 10 seconds.
Swarm of Locusts and other research.
Volvo thinks that a swarm of locusts might be a better example than a school of fish. Certainly their brains are much simpler.
Nissan, however, has done some work based on actual fish behaviour after studying bumblebees as well.