news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

tesla completes most of its data labeling at its buffalo factory

core tips:

1. tesla employs many full-time labelers who earn about $20 (about 142 yuan) per hour.

2. their work is monotonous and they may have to mark lane lines and curbs for months on end.

3. annotators can peek into the lives of tesla owners, which makes them feel a little "weird".

4. tesla uses two systems to monitor labelers and track their working hours, causing some people to be afraid to even go to the toilet. those who fail to meet the standards will be fired.

for tesla cars, when should the car start braking at a stop sign when it snows? when should the turn signal be turned on? how to distinguish between a traffic light and a full moon?

these are just some of the issues tesla's autopilot team deals with on a daily basis. tesla's assisted driving system relies on an army of data annotators who review thousands of hours of video from tesla owners and the company's internal test drivers. these annotators gradually teach tesla's artificial intelligence (ai) how to drive like a human driver, one 30-second video at a time.

tesla employs many labelers who work full-time at the company and earn about $20 (about 142 yuan) per hour.

business insider spoke to 17 current and former employees of tesla's data annotation team to get a glimpse into the work of the team behind the scenes that powers tesla's autopilot and full self-driving (fsd) systems. the team is based in three different tesla facilities: buffalo, new york; palo alto, california; and draper, utah.

the team works on projects with varying deadlines, from months to days, and with workflows that range from annotating short video clips to tracking still images or overlaying satellite data.

“the work can be monotonous at times,” said one former employee. “you might be annotating lane lines and curbs in thousands of videos eight hours a day for months on end.”

peep into other people's lives

these video clips can provide a unique window into the daily lives of tesla owners. five annotators said that in one case, a project required them to label data obtained from some owners’ garages through tesla’s sentry mode feature.

another program, called selfie, requires some annotators to annotate data captured by cameras inside tesla cars, according to two annotators. four other annotators said they were aware of the program, which is designed to teach tesla's system how to recognize when a driver is not paying attention to the road while using autopilot.

tesla says in its owner's manual that in-car cameras "share short video clips with tesla to help us develop future safety enhancements and continually improve the intelligence of features that rely on in-car cameras." tesla says owners must first opt ​​in to share their data so that tesla annotators can access the videos.

tesla assisted driving system

in other cases, annotators found themselves annotating data related to the travel routes of youtube celebrities and even elon musk himself, business insider previously reported.

“it feels weird to be peering into other people’s lives this closely,” said one current tesla annotator. “it feels weird to watch someone else’s daily driving, but it’s also an important part of correcting and improving the program.”

fifteen annotators said the videos came from all over the united states, as well as parts of europe and south america. two annotators recalled annotating videos from ukrainian car owners just as russia launched its military campaign against ukraine.

business insider reached out to tesla, musk and his legal team for comment but did not receive a response before press time.

annotators may encounter data from any number of countries in a single workflow, which means they must continually learn about different road rules in each region. seven former and current annotators said tesla sometimes seemed to take a more relaxed approach to those rules. for example, some annotators said they were told to ignore signs that said “no turn on red” or “no u-turn,” meaning they wouldn’t train the system to obey them.

“it’s an owner-first mentality,” said one former annotator. “i think the idea is that we want to train the car to drive like a human, not like a robot that just follows the rules.”

sometimes, annotators are asked to annotate videos from car accident and near-accident scenes.seven annotators recalled annotating videos that included tesla accidents or accidents involving neighboring vehicles. four annotators said that in one case, an annotator shared a video among colleagues that showed a young boy on a bicycle being hit by a tesla, which became one of the many videos and memes they used to communicate.

reuters first reported on the bike video and potential privacy issues at tesla’s labeling sites last year. soon after the article was published, tesla began limiting labelers’ access to videos outside of their assigned projects and added watermarks to some videos and images to make it easy to track which employees were sharing images, nine labelers said.

employee monitoring system

tesla has set up a fairly strict employee monitoring system in the buffalo factory. eleven labelers told business insider that there is a row of surveillance cameras that overlook the entire work area.

tesla also uses two different software systems to closely monitor its employees.four annotators said one of the systems, called humans, was used to assess how long they should spend on each video clip. annotators who continue to work too long within the required time may receive a bad review or be placed on a performance improvement plan (pip). the software was originally designed to help u.s. air force pilots and also has the ability to track employees' eye movements and record audio, but it is unclear whether tesla uses the software to track employees' eye movements.

tesla also uses a time metric called "flide time" to track the time annotators are active on the annotation software, according to 17 annotators. it tracks the number of keystrokes annotator makes and the time spent on the open annotation software, but does not track the time annotator uses other tools on the computer. depending on the level of the annotator, they need to record 5-7.5 hours of flide time, which means they must stay active on the software for at least that long.

musk

six labelers said they could face disciplinary action even if their active time was five minutes less than the required time. if they failed to meet the target three times in six months, they could be fired. some labelers have tried to object to the assessment indicators set by tesla, but with little success.

in february 2023, some employees at tesla's buffalo plant tried to form a union. union organizers at the buffalo plant told bloomberg that tesla tracks how long employees spend on each task and how long they are actively working each day by monitoring keystrokes, which has caused some people to be afraid to go to the bathroom. "people are tired of being treated like robots," said al celli, a member of the union's organizing committee.

in the same month, tesla fired dozens of workers at its buffalo plant. at the time, the national labor relations board (nlrb) filed a lawsuit accusing tesla of illegally firing some employees "in retaliation for union activity and to discourage union activity." however, tesla denied the allegation, saying that the employees were fired because of poor performance. the nlrb has not yet responded to the current progress of the lawsuit.

when tesla began building its assisted-driving program in 2016, reuters reported.the company outsourced the data labeling to a california-based company with offices in kenya, but tesla brought the project back in-house in 2019.

most recently, tesla's autopilot team was impacted by company-wide layoffs in april, with tesla laying off nearly 300 people in buffalo, pursuant to a worker adjustment and retraining notification (warn) act notice.

tesla has said its neural network will one day be able to train itself, but for now, it relies on human effort.

the work is essential to musk's vision for tesla, and over the years he has repeatedly stressed the importance of tesla's efforts to achieve self-driving. in 2022, musk said the difference between tesla being worth a lot of money or essentially nothing is self-driving technology.

tesla plans to launch a self-driving taxi service later this year, which is expected to be built on the company's same self-driving software and, of course, the lengthy, shot-by-shot analysis of video footage by its annotators.

for more first-hand news, please download the phoenix news client and subscribe to phoenix technology. if you want to read in-depth reports, please search "phoenix technology" on wechat.