Tesla admits remote drivers grab the wheel of its 'robotaxis'
The cars that glide through Austin and Phoenix with zero hands on the wheel are not always steering themselves. Tesla quietly told U.S. lawmakers that when its so-called robotaxis freeze in a maze of cones or a sudden white-out, a human in some distant office snaps into the driver’s seat—thousands of miles away.
Wired published the first pages of the company’s own submission last week. The phrase that jumps out is “remote operator intervention.” Translation: every time the neural net chokes on an unmapped construction zone, a gig-worker with a gamepad finishes the trip. The customer still sees an empty driver’s seat and pays for “Full Self-Driving.”
The autonomy mirage looks familiar
Roboticists roll their eyes. They have watched the same choreography with bipedal bots that dance on stage, then stumble once the velvet rope drops. Both fleets run on identical ingredients—lidar point clouds, imitation-learning stacks, reinforcement loops—yet the real world keeps serving edge cases no lab can afford to simulate. Snow flurries. A homeless man pushing a shopping cart against traffic. A freshly painted bike lane that is still wet.
Tesla’s filing lists 1,600 such “difficult scenarios” per month across its 500-vehicle pilot. Divide the numbers and you get three panic calls per car. That is not edge; it is routine.
Engineers inside the program say the remote stations resemble air-traffic control towers: three monitors, force-feedback steering, 5-millisecond latency fiber. One operator can babysit up to nine vehicles at once, clicking a priority queue that flashes red when the probability of disengagement tops 85 %. The hand-off takes 400 ms—faster than the average Uber driver’s reaction—but the philosophical damage is done.
Regulators smell bait-and-switch
The California DMV opened a fresh probe the morning after the documents leaked. Federal attorneys are asking whether marketing materials that promise “no driver needed” already violate the FTC Act. Investors yawned at first—Tesla stock dipped 2 % then rebounded—yet the liability iceberg is surfacing. If a remote operator misjudges a left turn, who carries insurance: Tesla, the contractor, or the absent owner sipping a latte in the back?
Meanwhile, competitors rush to differentiate. Waymo scrubs the phrase “fully autonomous” from its new rider app and replaces it with “driverless within defined conditions.” Cruise re-hired 200 safety drivers it laid off in December. Even Zoox quietly extends its geofenced test zone rather than expanding it.
The wider lesson is older than the combustion engine: demos seduce, but revenue demands robustness. Tesla booked $276 million in deferred FSD revenue last quarter—cash collected for a capability that, by its own paperwork, still needs ghost hands on the wheel. The company insists remote assistance is a temporary bridge. Skeptics note that bridge has lengthened every year since 2016.
Customers who paid $15,000 for the software upgrade feel conned. Reddit forums overflow with dash-cam clips titled “Look, no one—oh, wait, someone.” One user super-glued a Lego minifigure to the empty seat and captioned it “Tesla’s secret driver.” The joke stings because it is true.
Wall Street analysts keep raising price targets, betting the brand can outrun the lie. Yet the arithmetic of scale is merciless. If Tesla deploys 30,000 robotaxis by 2025, as Musk forecasts, maintaining three human supervisors for every nine cars implies a hidden payroll of 10,000 remote workers—an entire call center of drivers paid to pretend they are not driving.
The silence from Austin is louder than any earnings call. Tesla’s next master plan is due this fall; investors want a chapter titled “How we fire the humans.” Unless the chapter ends with a working silicon driver, the revolution will keep commuting on fiber-optic cables and underpaid gamers wearing headsets in Manila.
The steering wheel, it turns out, never left human hands—it just moved to a basement 8,000 miles away.