Justin Sullivan/Getty Images
- Robotaxis rely on large amounts of real-world and simulated data to improve their driving.
- Behind the scenes, human workers help to clean up and label the training data.
- Insiders estimate that the AV-support workforce is less than a couple of thousand people.
If robots take over the world, we'll remember the humans who helped them get there. Robotaxis are no exception.
They're not just engineers who have built the AI systems that drive the Waymos and Zooxes on the roads: Every day, thousands of people around the world sift through large amounts of driving data collected by cars retrofitted with clunky sensors.
The workers have different names — validators, annotators, labelers — but their objective is the same: help the AI driver figure out what it's seeing.
"What they're basically doing is helping the car understand where it is in space and time, and importantly helping the model to understand how it should safely navigate whatever scenario," Rowan Stone, CEO of Sapien, a data foundry that has clients like Zoox, told Business Insider.
The job can be as simple as helping the AI identify objects spotted on the roads through the sensors, be they cameras or lidars: Is that a cone? A stop sign? Tumbleweed?
Stone also pointed to scenarios like a police scene blocking the roads or a school bus dropping off kids — real-world situations that Waymo's robotaxis have struggled with — and said labelers provide more guidance on how to appropriately respond.
"Clearly that's where you need to bring humans back in," Stone said. "We need to re-hone the dataset, we need to use additional context to retrain the model, deploy your fix, and away you go."
A niche job in a big industry
The data labelling industry as a whole can be massive. Stone said Sapien has more than a million "contributors" around the world.
For AV systems in particular, that number is much smaller. Stone pegs the head count for autonomous vehicle-related operations at under 5,000 people worldwide. That number could scale as more robotaxis emerge.
TaskUs
Omar Zoubi, a VP at TaskUs, which provides third-party data labeling and remote support agents for companies like Waymo, told Business Insider that the company had just under 2,000 workers across its entire AV-related operations, which could double by the second quarter of this year.
Labeling itself may not be a very glamorous job: At Sapien, Stone said the average hourly rate is often set by customers or AV operators and can range from $3 to $6 per hour. TaskUs did not disclose how much its data labelers are paid.
The Sapien CEO said many of his company's contributors are based in Germany, Japan, and Southeast Asia. Overall, Sapien's "contributor" base spans about 100 countries, he said.
AI is taking some of the work
Artificial intelligence is also doing some of the legwork of data labeling, but it hasn't yet made the human's job obsolete.
Lukas Grapentine, a solutions engineering director for Sapien, told Business Insider that AI is "pre-labelling" some of the raw dataset and that the job of the humans is to check the AI's work.
When dealing with autonomous vehicle systems in which human lives are at stake, ensuring the AI is accurate is especially critical.
"It all comes down to being able to trust that data," Grapentine said.
Zoubi of TaskUs said he expects the role of the data labeler to evolve as automakers and AV companies get their hands on more data, which means coming across more complex driving scenarios. AI may be able to handle simpler tasks, but humans will have to take over interpreting more complex scenarios, he said.
"That's where I believe, at least my personal view on it, is where you'll see things shift — instead of doing just basic annotation and labeling of data, it'll be a lot more root-causing and fine-tuning that data to help the AVs operate and navigate those specific situations," Zoubi said.
Stone sees a similar projection — one in which AI models improve and rely on fewer humans over time as robotaxis get better at adapting to new cities and their quirks.
"I think the need for humans will trend down," he said, "but I don't think it will trend to zero."