How microtasking is fueling a surge in AI growth for self-driving vehicles

Self-driving cars have been the talk of the automotive industry for years, but we're feeling the push now more than ever. Is microtasking the secret to developing mainstream self-driving vehicles?

Google self driving car
Martyn Williams/IDGNS

People think of autonomous vehicles as incredibly powerful, automatic machines, but the software that guides those machines is programmed by individual people in a manual effort. We’re spoiled with technology that seems borderline magical, but even the AI programs that manage to learn on their own started out in the hands of human beings.

This manual effort is part of what’s slowing down the progress of self-driving cars (along with slow regulatory progress and logistical hurdles), but a new approach—microtasking—might offer a solution to the problem.

The programming challenges of self-driving vehicles

First, let’s focus on the manual programming challenges of self-driving vehicles. The basic architecture of the programs used for autonomous vehicles aren’t ridiculously complicated; in fact, they’re mostly based on traffic laws. For example, you’ll need to teach a machine that red lights mean “stop” and green lights mean “go,” and that dotted lines indicate things called “lanes.”

Another major challenge is teaching the AI complex decision-making, which is both programmatically and philosophically complicated; advanced engineers need to help the program understand what to do in complex situations, such as if a collision is unavoidable, or if a pedestrian unexpectedly jumps in front of the car.

The middle group of tasks is programmatically complicated, but isn’t hard for human minds to understand; in fact, we do it all the time. We need to teach programs how to recognize the concepts we’ve programmed into it.

For example, assuming we have sufficiently advanced visual sensors, how can we teach a program what counts as a dotted line and what counts as a coincidental line-like pattern in the nearby rocks? How can we teach a program to easily distinguish between real human beings and human-like shapes in its two-dimensional environment?

The solution for AI is similar to the solution for any human being; the key to learning is practice and repetition. The goal is to show a program instances of road signs, markings, and objects, and either instruct it or correct it until it learns for itself, highly reliably, to distinguish between them.

The problem is, because these markings vary from place to place, and conditions like weather and lighting can make things even more complicated, it takes millions of lessons to be sufficient.

Enter microtasking

This is where microtasking comes in. Microtasking isn’t entirely new—it’s a concept that’s been applied to teach AI systems before. But now that it’s being applied to autonomous vehicle technology, it could radically accelerate the growth of the industry.

Microtasking works by assigning thousands of people (or more) microtasks, which individually take mere minutes to complete, but collectively result in hundreds of person-hours of work being done.

For example, if 10,000 people each show an AI system what counts as a stop sign, and/or correct it when it guesses wrong a total of 10 times, you have 100,000 lessons completed in an hour or less, which is something even the fastest computer programmers wouldn’t be able to touch.

Accordingly, autonomous vehicle engineers are relying on microtasking to handle some of their most significant labor-intensive hurdles. If implemented correctly, it could shave thousands of hours of programming, testing, and finding mistakes off of major tech companies’ timetables for releasing autonomous vehicles to the public.

Disadvantages of microtasking

Despite the time-saving advantages of microtasking, there are still a few problems with the setup.


First, these people aren’t volunteering their efforts. Using a crowdsourced work platform like Mechanical Turk can lower costs for microtasks significantly, but if you’re recruiting hundreds of thousands of people, you’re still going to pay hundreds of thousands of dollars—or more.


Next, since you’re relying on human beings to make these assessments—human beings who may or may not have a vested interest in your success—there’s no guarantee of accuracy in their work. They may disregard some of the instructions, or even submit false information intentionally if they’re unsatisfied with the work.


Finally, not all people will recognize signs and markers the same way. This can result in mixed signals that make the AI’s job even more complicated.

The new paradigm

If the prospect of microtasking accelerating the development of self-driving vehicles has you excited, you might want to slow down for a second—it’s not time to junk your old vehicle just yet. Even with microtasking in place, working at maximum efficiency, there are still many hurdles to clear before Stage 4 or Stage 5 autonomous vehicles are made available to the public, including regulatory mandates, visual sensors, and the logistics of distribution. It may be three years, five years, or a decade before fully autonomous vehicles are in our hands, but we’re closing the distance faster than ever before.

This article is published as part of the IDG Contributor Network. Want to Join?