9
$\begingroup$

How are autonomous cars related to artificial intelligence? I would presume that artificial intelligence is when we are able to copy the human state of mind and perform tasks in the same way. But isn't an autonomous car just rule-based machines that operates due to its environment? They are not self-aware, and they cannot choose a good way to act in a never before experienced situation.

I know that many people often mention autonomous cars when speaking about AI, but I am not really convinced that these are related. Either I have a too strict understanding of what AI is or

$\endgroup$
4
$\begingroup$

There is a neat definition of artificial intelligence, which circumvents the problem of defining "intelligence" and which I would ascribe to McCarthy, the founder of the field, although I can only find it now in this book by H. Simon:

"… having to do with finding ways to do intelligent tasks, to do tasks which, if they were done by human beings, would call for our human intelligence."

So, at its core we call the automation of every task AI, that can only be done by the human mind. At the time people thought that a computer able to play chess would also be intelligent in other ways. When this turned out to be false, the term AI was split into "narrow or weak AI", i.e. a program able to do one task of the human mind, and "general or strong AI", a program that can do all the tasks of the human mind.

Self-driving cars are narrow AI.

Note, that all these definitions don't specify whether these programs copy the way the human mind works or whether they come to the same result via completely different algorithms.

$\endgroup$
  • 1
    $\begingroup$ It may be that self-driving cars that can handle a variety of changing conditions (broken traffic lights, police directing traffic, poorly marked detours) on not-very-well-mapped roads requires fairly strong AI. It hasn't been done yet and it's not clear that it can be done with anything like current technology. $\endgroup$ – antlersoft Oct 12 '16 at 16:56
  • $\begingroup$ @antlersoft Yes, in envs with more mixed use where understanding language and eye contact is a key to driving, it's more or less strong AI. A researcher from Uber's autonomous lab told me that driving in most other countries is closer to cycling in the US than to driving in the US. $\endgroup$ – Adam Bittlingmayer Mar 24 '17 at 12:48
1
$\begingroup$

Other answers tell about sets of instructions for the car in certain situations, or a goal seeking machine, while in fact, self-driving cars don't have a specific set of instructions. Most self-driving cars use deep learning to figure out what to do at certain events. We don't tell them what to do. They learn what to do by example.

The neural networks used to automate cars need massive amounts of data to train. Using the data, the car can figure out what the best action is for certain events.

According to this video Tesla's Autopilot had only one casualty in 300.000.000 miles. For human drivers, the number of casualties in 2014 was 32.675. That is per 300.000.000.000 miles. That means 1 in 90 million human drivers cause a fatal accident, compared to 1 in 300 million for automated cars. Deep Learning surpassed our own 'safety-rate', not by instruction, but by learning what to do itself. If that isn't AI, I don't know what is.

$\endgroup$
0
$\begingroup$

Self driving cars exhibit a level of agency and multi-domain resilience. By certain definitions they are self aware and they are definitely designed to fail safely in a large number of potentially unknown circumstances, which is similar to biological agents.

AI really has to do with the study of non-biological agents and their methods of agency. Everything else is just computer science, algorithmic efficiency, biology, art, etc. Eventually the study of biological and non-biological agency will converge, though, and we'll just call it the study of "intelligence."

$\endgroup$
0
$\begingroup$

Others have given very detailed answers, this is my layman view of the problem statement. The self driving car is a 'goal seeking' machine. It has a set of goals with different priorities. Example. Safety of Occupants, Safety of others, Go from Point A to Point B etc. Some are negotiable, other not so.

To satisfy the goals, the system should use the inputs available (radar, GPS, Camera etc) to determine what is the best possible course of action. At times when it doesn't have all the info (a truck which is hiding a speed sign), it still has to take a decision (historic memory or through awareness of its surroundings) to satisfy its design goals. Hence the AI.

$\endgroup$
0
$\begingroup$

Autonomous vehicles are dependent upon AI technology in that, to be autonomous in their driving or piloting, they cannot be controlled by people. Therefore they must make complex decisions required of drivers and pilots at least as safely and reliably as human drivers or pilots.

  • They must recognize objects to the degree that both the value and the typical behavior can be assigned to those objects (i.e. people, pets, property, barriers, curbs, grass, trees, bridges)
  • They must map trajectories of a wide array of object types based on their object type, what is known about that type of object, detectable variations such as age or condition, and what the object appears to be involved in doing at the time.
  • They must be able to acquire publicly available representations of drive-able roads (route segments, connection points, and other data), match the representation with the current state of the roads, and track their progress along an intended route to the destination.
  • They must plan their course in lieu of these real time and difficult to predict actions, traffic law, traffic conventions, traffic signs and signals, given destination, known possible routes, discontinuities, and anomalies.
  • They must be able to alter the plan to reach the destination if at all possible regardless of changes and challenges encountered.

Driving or piloting a vehicle is an intelligence intensive task. The only reason AVs will likely surpass human driven vehicles on the road in the near future in terms of the distributions of rates of fatalities and injuries per million meter of travel in the near future is because humans have two key handicaps that offset their intelligence potential as drivers.

  • Carelessness, as defined as multitasking either mentally or physically at a time when hazards might appear
  • Selfishness, as defined as risking the life, health, or property of others to gain a transportation related or psychologically related advantage

Although the above two appear to be subjective, they can be easily proven empirically by taking a sample of traffic patterns at any point in time in any highly trafficked road in the world. This is less true of pilots.

We should not presume that artificial intelligence in AVs is achieved when the behavior of the human mind is copied. That is the criteria for Alan Turing's Imitation Game, a test that was intended to define intelligence in the context of natural language dialog. But words don't normally kill people directly. Vehicles often do.

It would be a very limited vision the potential AV design space to consider human minds as the model of driving excellence. The tasks should not be performed in the same way by the AI system. The AI design objectives of AVs should be more consistent with these concerns and interests.

  • Road or sky safety laws
  • Ethics regarding right of way in normal and emergency situations
  • Civil rights concerns in terms of equal access to public resources
  • Balancing of spacial flow details to maximize transportation throughput
  • Collision aversion when difficult to predict risks emerge

These requirements on the cognitive and adaptive capabilities of the driving or piloting AI are not solely rule-based and mechanical. The vehicle itself is mostly mechanical in its operation, but it too presents surprises like blowouts or other difficult to predict failures. Vehicle control is not at all like chess or a game with a fixed rules of play and fixed game-play environment.

Although the intelligence requirements do NOT include self-awareness of itself as an intelligent system, there are forms of self-awareness required.

  • The relative position of the exterior surface of the vehicle and its projected path relative to that of other objects
  • The condition of the operational parts of the vehicle
  • The mass and location of passengers and any other transported objects in the vehicle

The question ended with an interesting and challenging requirement.

Choose a good way to act in a never before experienced situation

That is perhaps the most challenging aspect of AV driving or piloting system design.

Returning to the question of, "Why are autonomous cars categorized as AI?", the meaning of AI is indeed a critical aspect of answering well. Taken literally, the term artificial intelligence specifies two things.

  • It is artificial, in that it does not naturally occur in nature
  • It is intelligent, in that it adapts in ways that, if those ways are mechanical, they are mechanical at a level of detail that is beyond obviousness without considerable study

As year dependent and culturally dependent as that definition of intelligence is, no other definition is quite as sustainable over decades from both scientific and linguistic perspectives. By narrower definitions, AVs may not require AI, but there is no compelling scientific reason to narrow the definition of AI to a subset of this previous definition.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.