In August, addressing Bloomberg the celebrity of artificial intelligence, Andrew Ng, said the fastest way to create reliable autonomous vehicles is ] and not cars. "What we are telling people is," Please, be legal and considerate, "Ng told Bloomberg.
Ng's remarks, which come at a particularly sensitive moment in the brief history of driverless cars, have caused an uproar in the AI community, sparking criticism and approval from various experts .
In recent months, autonomous cars have been involved in several incidents. One of them resulted in the death of the pedestrian .
Most researchers and AI specialists agree that driverless cars have still not made enough progress to let them roam the streets without a redundant supervised driver and be ready to jump to the wheel in case of a problem.
But this is the end of the agreements. There is a big gap between when driverless cars will be ready for the road, the transition phase and how to meet the challenges of autonomous driving.
How do autonomous cars understand the world around them
For vehicles to drive alone, they must understand the world around them as (or better) human drivers, so they can move around the streets, stop at stop signs and at traffic lights and avoid obstacles. like other cars and pedestrians.
Modern computer vision has come a long way thanks to the progress of in-depth learning which allows it to recognize different objects in images by examining and comparing millions of examples and cleaning the visual. templates that define each object. Although particularly effective for classification tasks, in-depth learning suffers from severe limitations and may fail unpredictably.
This means that your driverless car could hit a truck in broad daylight or worse, accidentally hit a pedestrian. The computer vision technology currently used in autonomous vehicles is also vulnerable to conflicting attacks, in which hackers manipulate the artificial intelligence input channels to force them to make mistakes.
For example, researchers have shown that they could fool an autonomous car to avoid recognizing the stop signs in sticking black and white labels.
One day, AI and computer vision might become good enough to avoid committing the erratic errors that driverless cars currently make. But we do not know when it will happen and the industry is divided on what to do until then.
Improvement of computer vision technology of driverless cars
Tesla the company founded by the eccentric Elon Musk, thinks to be able to surpass the limits of the artificial intelligence that powers autonomous vehicles by sending more and more data. This is based on the general rule that the more you provide deep learning algorithms for quality data, the better they become.
Tesla has equipped its vehicles with a set of sensors and collects as much data as possible from these sensors. These data allow the company to continuously train its artificial intelligence with the data that it has collected from the hundreds of thousands of Tesla cars that roam the streets of different parts of the world.
The reasoning is that, as its artificial intelligence improves, Tesla will be able to deploy new updates of all its vehicles and make them more efficient in the performance of their autonomous driving functions. The advantage of this model is that it can all be packed in a mainstream vehicle. It does not need expensive additional equipment attached to the car.
To be fair, it is a model that only a company such as Tesla can execute. Like many other things, automobiles go through a period of transition with the calculation of and connectivity becomes ubiquitous . In this regard, Tesla is more advanced than other companies, because instead of being a car manufacturer that tries to adapt to new technological trends, it is a technology company that manufactures cars.
Tesla's cars are actually wheel-based computers and can be constantly upgraded through live software updates, a feat harder to exploit for other companies .
This means that Tesla will be able to gradually improve the driving capabilities of its vehicles by collecting more data and continuing to train its models to improve its AI.
Tesla also has the opportunity to train his AI by "driving in the shadows", passively watching the driver's decisions and weighing him in the way he would act if he was in a similar situation in driving mode autonomous.
All of this works as long as the problem of computer vision is a problem that can be solved with more data and better training. Some scientists believe that one must think of AI technologies beyond deep learning and neural networks . In this case, Tesla will have to restructure the specialized computer hardware supporting the self-driving features of its vehicles.
Equipping autonomous cars with complementary technologies
Google and Uber two other companies that have invested heavily in autonomous driving technologies, have leveraged several technologies to offset the shortcomings of the artificial vision of driverless cars. The most important of them is light detection and telemetry (lidar).
Lidar is an evolving field and various companies use different technologies to perform its functions. Lidar patents and intellectual properties were at the center of a long legal battle between Google and Uber which was settled at 245 million dollars earlier this year.
In summary, lidar works by sending millions of laser pulses in slightly different directions to create a 3D representation of the area surrounding the car as a function of the time it takes for the pulses to strike an object and return. This is the rotary cylinder that you see above some autonomous cars (all lidars do not look like this, but it has somehow become an icon of the industry).
In addition to lidar, these companies also use a radar to detect different objects around the car and evaluate traffic and road conditions. The following video shows how the technology works.
The addition of all these technologies certainly makes these vehicles much better equipped than Tesla's only visual approach. However, that does not make their technology perfect. In fact, an accident that made headlines earlier this year is about an Uber vehicle that was in autonomous driving mode.
In addition, the approach of Google and Uber makes the deployment of driverless cars on roads much more expensive and more difficult. Google and Uber have traveled millions of kilometers with their autonomous technology and have collected a lot of road data, but that does not start to compete with the amount of data that hundreds of thousands of vehicles Tesla sold collect. In addition, the addition of all this equipment to a car is expensive.
Lidars alone add between $ 7,000 and $ 85,000 to the cost of a car, and their form factor is not very appealing. Add to that the costs of all other sensors and equipment to be added to the vehicle's post-production, and you could double or triple the cost of your car .
If scientists can decipher the code of computer vision and create an artificial intelligence capable of understanding the surrounding world and human factors, then Tesla will be the winner of the race. It will be necessary to launch a new update and all its cars will become magically able to drive in an almost perfect autonomous way.
On the other hand, if the current trends of narrow AI never reach to equal the human drivers, Google and Uber will be the winners – that is, they can reduce the costs of lidar and other driverless car equipment. Automakers could then consider equipping their vehicles with autonomous driving technology without significantly increasing costs.
Advancing autonomous driving by repairing pedestrians
Andrew Ng is part of a handful of Amnesty International's opinion leaders who think that shortening our path to autonomous driving is to prevent pedestrians from causing unannounced behavior of driverless cars .
This basically means that if you are jaywalking and an autonomous vehicle strikes you, it's your fault. To the extreme, this would practically make cars in trains, where pedestrians are responsible for everything that happens to them if they stand on the railroad.
The establishment of a strict rule of conduct for pedestrians and the limitation of their movement on the roads will certainly make the environment much more predictable and navigable for autonomous cars.
But not everyone is convinced by this proposal and many of them are questioning it, including Professor Gary Marcus of New York University, who says that changing human behavior will only make that "move the goals".
Rodney Brooks, another legend of AI and robotics, rejects Ng's proposal . "The great promise of self-driving cars has been to eliminate the dead from the road," he said, adding that Ng asserted "that they will eliminate the dead from the road as long as all humans will be trained to change their behavior? " We also think that we will change human behavior so easily that we would not need autonomous cars to eliminate road accidents.
But Ng does not think that moving the poles is an absurd idea, arguing that humans tend to adapt to new technologies, as they did with the railways. The same thing can happen with driverless cars.
Be that as it may, a compromise between fully intelligent cars that can meet all possible scenarios (for example, a pedestrian jumping in the middle of the street with a pogo stick) and a rail-style frame where pedestrians are totally prohibited Moving to areas with self-driving vehicles will likely help ease the transition as technology develops further and autonomous cars become the norm.
Adaptation of urban infrastructure to autonomous cars
Another solution to the challenge of driverless cars is to repair the roads and environments in which they operate. This operation also has a precedent.
For example, with the advent of automobiles, roads were improved and created for very high speed tires. With the advent of aircraft, airports were created. In cities where bikes are very popular, separate lanes have been created for bicycles.
What is the infrastructure of driverless vehicles? Academics from the Edinburgh Business School propose in an article from Harvard Business Review to create intelligent environments for autonomous cars.
Currently, driverless cars have no way of interacting with their environment and all they learn comes from their sensors, their lidars, their radars and their video streams. . By incorporating elements of the Internet of Things (IoT) into the roads, bridges and other elements of the infrastructures of the city, we can make them more understandable for autonomous cars.
For example, installing sensors at specific intervals on the sides or in the middle of roads can help driverless cars locate their limits, whether the road is clear or covered with snow or mud or It is buried under two inches of floodwater.
Sensors can also provide autonomous cars with information on road and weather conditions, such as their sliding and careful driving.
Driverless cars must also be able to perform machine-to-machine (M2M) communications with other manual or stand-alone vehicles located nearby. This will help them coordinate their movements and avoid collisions more accurately.
One of the challenges of this model is that vehicles live for decades. This means that today's cars will remain on the road in the 2030s. So you can not expect every vehicle to be equipped with sensors and M2M capabilities. Moreover, we can not expect all roads in the world to suddenly develop into intelligent sensors.
However, driverless cars, whose number is currently very limited, can be equipped with technology to search for nearby smart sensors and, where appropriate, interact with them to offer a more safe. And if they can not find a standard smart sensor in their environment, they can use their own local equipment to navigate their environment.
When will driverless cars become the norm?
There are different estimates of the time it takes for driverless cars to drive around the streets with manual and semi-autonomous vehicles. But it has become clear that meeting the challenges is much more difficult than we first thought.
Our cars may one day become smart enough to cope with all possible scenarios. But this will not happen overnight and it will probably take several steps and steps at different levels. In the meantime, we need technologies and practices that will facilitate the transition until we can have autonomous vehicles capable of making our roads safer, our cities cleaner and our travel less expensive.