Inside ‘Project Rodeo,’ the Tesla effort pushing the limits of self-driving technology
Test drivers said they sometimes navigated perilous scenarios, particularly those drivers on Project Rodeo’s “critical intervention” team, who say they’re trained to wait as long as possible before taking over the car’s controls.
Since 2013, Elon Musk has promised that Tesla will have a self-driving car. To get there, the company has leaned on a specialized group of test drivers who are part of what’s known internally as “Project Rodeo.”
Test drivers on Project Rodeo say they push the company’s self-driving software to its limit. They work to bridge the gap between driver-assist software and fully autonomous driving. Operating on open streets with other vehicles, cyclists, and pedestrians, they have tested unreleased software that will be crucial to Tesla’s push into autonomous driving.
Test drivers said they sometimes navigated perilous scenarios, particularly those drivers on Project Rodeo’s “critical intervention” team, who say they’re trained to wait as long as possible before taking over the car’s controls. Tesla engineers say there’s a reason for this: The longer the car continues to drive itself, the more data they have to work with. Experts in self-driving tech and safety say this type of approach could speed up the software’s development but risks the safety of the test drivers and people on public roads.
“The idea is that you’re a cowboy on a bull and you’re just trying to hang on as long as you can,” a former test driver who trained for the critical-intervention team in San Francisco said.
Business Insider spoke with nine current and former Project Rodeo test drivers and three Autopilot engineers in states including California, Texas, and Florida. The drivers worked on training Tesla’s Full Self-Driving software and its Autopilot software, which, despite the products’ names, require a licensed driver at the wheel. Most asked to remain anonymous, citing a fear of professional reprisal, but their identities are known to Business Insider. Eight of the drivers described experiences that occurred over the past year, mostly between November and April.
Test drivers on Project Rodeo say they push the company’s self-driving software to its limit.
None of the test drivers who spoke with BI said they had been involved in a crash.
Five who worked for the company in 2024 said they narrowly avoided collisions, including almost hitting a group of pedestrians. One former critical-intervention driver in Texas told BI that they sometimes ventured into their city’s bar district late at night to see how Tesla’s FSD software reacted to drunk patrons spilling out after last call. The former driver in San Francisco recalled riding around Stanford University during training, testing how close FSD would allow the vehicle to get to people at crosswalks before they had to take over. And a third critical-intervention driver said they allowed the car to speed through yellow lights and drive 35 mph under the speed limit on an expressway to avoid disengaging the system.
A former Autopilot engineer told BI that while testing is done on open roads, Tesla runs hundreds of simulations and sometimes tests difficult scenarios on a closed course before rolling out new software to test drivers’ vehicles.
Tesla did not respond to a detailed list of questions about Project Rodeo and its self-driving technology.
The test drivers’ experiences highlight the balancing act Tesla and other automakers face as they prepare their self-driving software for widespread consumer use.
Experts say public testing is crucial and can help identify safety issues before the technology hits the market. Missy Cummings, a former safety advisor for the National Highway Traffic Safety Administration, said that while practices vary, she believes many autonomous-vehicle companies likely employ tactics similar to Tesla’s.
“In theory, these drivers have gone through training, and eventually these cars do need to be able to operate in the public domain,” Cummings said. She added that clearly marked vehicles could help the public better identify test drivers.
Safety experts say fragmented and limited autonomous-vehicle regulations, coupled with self-reporting by automakers, create a complex environment where companies balance public safety with getting their products ready for commercial use.
“There are very few rules around autonomous testing and a lot of dependency on self-reporting,” said Mark Rosekind, a former NHTSA administrator and chief safety innovation officer for Zoox, an Amazon-owned autonomous-taxi firm. “If companies aren’t reporting, it’s hard to know what’s going on.”
In the past decade, authorities have investigated several automakers, including Tesla, Waymo, and Cruise, over crashes involving self-driving or driver-assist software.
Tesla co-founder and CEO Elon Musk has said that self-driving is “really the difference between Tesla being worth a lot of money or worth basically zero.”
A lot depends on Tesla’s promise of autonomous driving; Musk said in 2022 that self-driving was “really the difference between Tesla being worth a lot of money or worth basically zero.” The Morgan Stanley analyst Adam Jonas wrote in a note last month that the company’s future valuation was “highly dependent on its ability to develop, manufacture, and commercialize autonomous technologies.” Tesla’s stock fell by 10% on October 11, the day after the company’s robotaxi event. It rebounded after Tesla reported earnings on October 23, and it’s up by more than 3% overall in the year to date.
Inside the day-to-day of a Project Rodeo driver
Tesla went on a hiring spree this year, bringing on test drivers in at least half a dozen US cities, a review of LinkedIn profiles suggests.
A job listing from 2023 said test drivers needed a “clean driving record, safe driving habits, and a minimum of 4 years of licensed driving experiences.” Eight drivers told BI that the onboarding process included two to three weeks of hands-on training, including test drives with a trainer in the passenger seat.
According to employees and internal documents, one specialty within Project Rodeo works to replicate the job of a ride-hailing driver by picking random points on a map and driving between them. Those working on a different team, known as the “golden manual” team, drive manually, without any assistance, to train FSD software on what by-the-book, error-free driving looks like.
Critical-intervention test drivers, who are among Project Rodeo’s most experienced, let the software continue driving even after it makes a mistake. They’re trained to stage “interventions” — taking manual control of the car — only to prevent a crash, said the three critical-intervention drivers and five other drivers familiar with the team’s mission. Drivers on the team and internal documents say that cars rolled through red lights, swerved into other lanes, or failed to follow posted speed limits while FSD was engaged. The drivers said they allowed FSD to remain in control during these incidents because supervisors encouraged them to try to avoid taking over.
The critical-intervention drivers recalled multiple instances where they felt unsafe but believed intervening could put their jobs at risk. The former driver in Texas recalled taking over only after FSD nearly rammed the car into the side of a vehicle stopped at an intersection.
Non-critical-intervention drivers said they also felt pressure to push the system as far as possible. Five current and former employees said that they were instructed to intervene if they became uncomfortable with the software’s behavior but that they sometimes received feedback from their supervisors if they were considered to have disengaged too early.
John Bernal, a former test driver and data analyst at Tesla, said test drivers dealt with risky situations as far back as 2022. (Bernal was terminated that year; he said he was fired for sharing videos on his YouTube channel that showed his personal Tesla malfunctioning while using FSD.) He described instances where he broke traffic laws in order to collect data. Bernal said that his supervisors never instructed him to break the law but that he sometimes felt it was the only way to get the data the company wanted.
He recalled one test in 2022 that was designed to see how well the system recognized a red light.
“My training was to wait until the wheels touched the white line before I could slam on the brakes,” Bernal said. He said he sometimes ended up in the middle of the intersection if the system didn’t work correctly.
He also worked to train the autonomous software on “vulnerable road users” — defined by the Department of Transportation as pedestrians, bicyclists, people on scooters or in wheelchairs, or highway workers on foot — when he manually drove the “Ground Truth Machine,” a Tesla outfitted with lidar and radar sensors to help the system map and identify objects.
“I’d drive over double lines to get close to a bike,” Bernal said. “I would go obnoxiously slow through an alleyway where drunk people were, and I would be extremely rude and get really close to people.”
“We want the data to know what led the car to that decision,” a former Autopilot engineer said.
Two years later, test drivers were asked to train the system near pedestrians, test drivers said. Five recalled a bug with FSD that made vehicles brake too early at crosswalks. To improve its performance, they were instructed to interact with pedestrians as frequently as possible.
Sometimes, the drivers said, the software would slam on the brakes when no one was at the crosswalk; other times, it wouldn’t stop at all. According to Tesla employees and internal documentation, FSD’s performance depended heavily on its software version, and the versions appeared to operate at different levels of caution.
The former San Francisco driver said that as they drove around Stanford University, their trainer, another test operator with more experience on the team, chastised them for braking too early. They recalled that at one point they came within three feet of hitting a bicyclist at a roundabout.
“I vividly remember this guy jumping off his bike. He was terrified,” the driver told BI. “The car lunged at him, and all I could do was stomp on the brakes.” They said the trainer was pleased by the incident. “He told me, ‘That was perfect.’ That was exactly what they wanted me to do.”
The driver added that “it felt like the goal was almost to simulate a hit-or-miss accident and then prevent it at the last second.”
The former Autopilot engineer said it was better for training to see whether the software could correct itself. They also said that not intervening when the car acted abnormally — including veering into another lane or doing something that confuses another driver — was important for training. Human motorists don’t always drive rationally, they explained, and the software needs to know how to respond. It’s also easier to parse the data if there are fewer driver interventions, they said.
“At the critical juncture where it’s about to make the key decision,” they said, it’s helpful to see whether the software makes the right or wrong call. “We want the data to know what led the car to that decision,” the engineer said. “If you keep intervening too early, we don’t really get to the exact moment where we’re like, OK, we understand what happened.”
A ‘Wild West’ with little regulation
Tesla is one of many automakers attempting to make autonomous vehicles a reality. Waymo, backed by Alphabet, launched the first driverless taxi service in Phoenix in 2020.
“In many ways, it’s like the Wild West out there,” said Cummings, the former NHTSA safety advisor. “There is very little regulation around training or informing the public about testing.”
The stakes of data collection on public roads are high. In 2018, a self-driving Uber with a person behind the wheel struck and killed a pedestrian in Arizona. Cruise paused testing after one of its driverless vehicles hit a pedestrian in October 2023. Another vehicle had already hit the pedestrian when the Cruise car struck her, dragging her 20 feet before stopping. It resumed testing with safety drivers in some cities in May.
A Waymo self-driving taxi stopped at a red light in Los Angeles, California, in March 2024.
Two former Waymo employees said that they had a team similar to Tesla’s critical-intervention team but that Waymo’s version of critical-intervention testing was limited to closed tracks with dummies. Two former Cruise employees said that they had mapping teams and teams that tested on closed courses and public roads but that, unlike at Tesla, those teams were instructed to take over as soon as the software went off track, and they typically tested with at least two people in the car.
A Waymo spokesperson said the company’s safety framework included rigorous testing in controlled environments and on public roads. A Cruise spokesperson said the company’s vehicles were designed as fully autonomous systems and were therefore fundamentally different from Tesla’s driver-assistance technologies.
Philip Koopman, an autonomous-driving expert at Carnegie Mellon University, said that Tesla’s critical-intervention approach, as described to him by BI, was “irresponsible” and that the company should be playing out all “critical scenarios” on a closed course.
“By allowing the software to continue misbehaving to the point a test driver needs to avoid a crash, Tesla would be imposing a risk on other road users who have not agreed to serve as test subjects,” Koopman said.
Alex Roy, a general partner at NIVC and a former director of operations at the autonomous-driving startup Argo AI, said companies should be correcting the software as soon as it strays from the course, particularly on public roads.
“You should play those mistakes out in a simulation, not on an open road,” Roy said. “Real-world testing is necessary, but real-world mistakes are not.”
“If you have a parent that’s holding the bike the entire time, it never gets to learn.”
A former Tesla Autopilot engineer
The former Tesla engineer said they doubted that computer simulations are sophisticated enough to replicate the data generated by real-world driving. The former engineer said that, to help the software improve, it was best for test drivers to avoid intervening whenever possible.
“If you have a parent that’s holding the bike the entire time, it never gets to learn,” the engineer said.
Test drivers on Project Rodeo felt this keenly.
“You’re pretty much running on adrenaline the entire eight-hour shift,” one former test driver in the Southwest said. “There’s this feelingTwo former Waymo employees said that they had a team similar to Tesla’s critical-intervention team but that Waymo’s version of critical-intervention testing was limited to closed tracks with dummies. Two former Cruise employees said that they had mapping teams and teams that tested on closed courses and public roads but that, unlike at Tesla, those teams were instructed to take over as soon as the software went off track, and they typically tested with at least two people in the car.
A Waymo spokesperson said the company’s safety framework included rigorous testing in controlled environments and on public roads. A Cruise spokesperson said the company’s vehicles were designed as fully autonomous systems and were therefore fundamentally different from Tesla’s driver-assistance technologies.
Philip Koopman, an autonomous-driving expert at Carnegie Mellon University, said that Tesla’s critical-intervention approach, as described to him by BI, was “irresponsible” and that the company should be playing out all “critical scenarios” on a closed course.
“By allowing the software to continue misbehaving to the point a test driver needs to avoid a crash, Tesla would be imposing a risk on other road users who have not agreed to serve as test subjects,” Koopman said.
Alex Roy, a general partner at NIVC and a former director of operations at the autonomous-driving startup Argo AI, said companies should be correcting the software as soon as it strays from the course, particularly on public roads.
“You should play those mistakes out in a simulation, not on an open road,” Roy said. “Real-world testing is necessary, but real-world mistakes are not.”
“If you have a parent that’s holding the bike the entire time, it never gets to learn.”
A former Tesla Autopilot engineer
The former Tesla engineer said they doubted that computer simulations are sophisticated enough to replicate the data generated by real-world driving. The former engineer said that, to help the software improve, it was best for test drivers to avoid intervening whenever possible.
“If you have a parent that’s holding the bike the entire time, it never gets to learn,” the engineer said.
Test drivers on Project Rodeo felt this keenly.
“You’re pretty much running on adrenaline the entire eight-hour shift,” one former test driver in the Southwest said. “There’s this feeling that you’re on the edge of something going seriously wrong.” that you’re on the edge of something going seriously wrong.”
