News
Cruise to reduce self-driving fleet by 50% in SF following robotaxi crash
The California Department of Motor Vehicles (DMV) has asked General Motors’ self-driving unit, Cruise, to temporarily reduce its robotaxi fleet by 50% in San Francisco. The request came after one of the company’s driverless vehicles was involved in a crash with a fire truck last week.
The California DMV noted that its is investigating “recent concerning incidents” involving autonomous robotaxis in the city. With this in mind, the DMV has advised that Cruise must reduce its San Francisco fleet by half until the investigation is completed and the self-driving company takes appropriate actions to improve road safety.
Following is the CA DMV’s statement about Cruise’s recent robotaxi incidents.
“Safety of the traveling public is the California DMV’s top priority. The primary focus of the DMV’s regulations is the safe operation of autonomous vehicles and safety of the public who share the road with these vehicles.
“The DMV is investigating recent concerning incidents involving Cruise vehicles in San Francisco. The DMV is in contact with Cruise and law enforcement officials to determine the facts and requested Cruise to immediately reduce its active fleet of operating vehicles by 50% until the investigation is complete and Cruise takes appropriate corrective actions to improve road safety. Cruise has agreed to a 50% reduction and will have no more than 50 driverless vehicles in operation during the day and 150 driverless vehicles in operation at night.
“The DMV reserves the right, following investigation of the facts, to suspend or revoke testing and/or deployment permits if there is determined to be an unreasonable risk to public safety.”
With Cruise agreeing to reduce its fleet by 50%, the company also released a blog post explaining the circumstances that led up to the accident. As per Cruise, the robotaxi identified the emergency vehicle almost immediately as it came into video. The robotaxi operator also noted that the vehicle initiated a braking maneuver and reduced its speed, but it was ultimately unable to avoid the collision.
All in all, the damage is super light. Good learning experience. A deaf, colorblind, human driver would have probably also gotten clipped.
In this situation, if you don't hear the truck and you don't notice the approaching scattered red light, reflecting on the buildings/street… pic.twitter.com/mTAGEQkAEk
— oFFMetaSweat (@oFFMetaSweat) August 18, 2023
Following is Cruise’s post about the incident.
“While our investigation is ongoing and we remain in contact with city officials and regulators, we wanted to provide an update on our preliminary analysis on the incident involving an emergency vehicle colliding with a Cruise AV.
“First and foremost, our primary concern remains with our passenger and their well-being. We have been in contact to offer support and will remain in touch.
“In terms of what occurred around the scene of the collision there are many aspects that looked typical from the AV’s perspective and several factors that added complexity to this specific incident.
“The AV positively identified the emergency vehicle almost immediately as it came into view, which is consistent with our underlying safety design and expectation. It is worth noting, however, that the confines of this specific intersection make visual identification more challenging – for humans and AVs alike – as it is significantly occluded by buildings, meaning that it is not possible to see objects around the corner until they are physically very close to the intersection.
“The AV’s ability to successfully chart the emergency vehicle’s path was complicated by the fact that the emergency vehicle was in the oncoming lane of traffic, which it had moved into to bypass the red light.
“Cruise AVs have the ability to detect emergency sirens, which increase their ability to operate safely around emergency vehicles and accompanying scenes. In this instance, the AV identified the siren as soon as it was distinguishable from the background noise.
“The Cruise AV did identify the risk of a collision and initiated a braking maneuver, reducing its speed, but was ultimately unable to avoid the collision.
“During the course of more than 3 million miles of fully autonomous driving in San Francisco we’ve seen an enormous number of emergency vehicles – more than 168,000 interactions just in the first 7 months of this year alone. Our first responders are trying to balance keeping all of us safe while quickly responding to emergency scenes and we’re grateful for their work and dedication.
“We realize that we’ll always encounter challenging situations, which is why continuous improvement is central to our work. We will continue to work in partnership with regulators and city departments on EMV interactions to reduce the likelihood of incidents like these happening again.”
Don’t hesitate to contact us with news tips. Just send a message to simon@teslarati.com to give us a heads-up.
Tesla has reached a notable benchmark in its autonomous driving program after its Robotaxi fleet surpassed one million miles of unsupervised operation. The company made the announcement during its Cybercab event in Austin on September 3.
Tesla Vice President of AI Ashok Elluswamy told attendees he was happy to report the fleet had achieved one million miles of unsupervised Robotaxi operation as a testament to safety.
The new total marked a sharp increase from the 380,000 unsupervised miles Tesla disclosed during its second-quarter 2026 earnings update in late July.
In roughly six weeks, the company added about 620,000 miles. That acceleration followed Tesla’s decision to remove in-vehicle safety monitors from most of its operations outside the San Francisco Bay Area.
Credit: Tesla
Tesla first launched Robotaxi service in Austin in June 2025 with safety drivers present. It later began fully unsupervised rides and expanded into Dallas, Houston, Miami, Orlando, and Tampa. The San Francisco Bay Area remains the exception, where a safety monitor still rides in the vehicle under California permitting rules.
The company has not released a city-by-city breakdown of the one million unsupervised miles.
The milestone arrived as Tesla began offering public Cybercab rides in Austin. The purpose-built vehicle has no steering wheel or pedals and is designed only for autonomous ride-hailing. Production versions joined the existing fleet of modified Tesla vehicles already operating in the service.
Tesla’s unsupervised mileage is growing at a double-digit weekly rate according to earlier company comments, yet its fleet size remains modest compared with established competitors. Waymo has accumulated more than 200 million fully autonomous rider-only miles. Tesla has described its own unsupervised operations as having recorded zero notable incidents in the period leading up to the July update.
The one-million-mile figure reflects Tesla’s shift from supervised testing to broader driverless service in multiple states. It also highlights the company’s strategy of using both existing Model Y vehicles and the new Cybercab to scale its network.
Credit: Tesla
Whether the rapid recent growth continues will depend on further city expansions, regulatory approvals, and the performance of the purpose-built Cybercab in everyday paid rides. Tesla has not specified how many of the latest miles involved the new vehicle versus the rest of the fleet.
The announcement underscores Tesla’s progress toward a larger robotaxi network while illustrating the remaining gap in total autonomous experience relative to longer-operating rivals.
Tesla AI lead Ashok Elluswamy said this week that 24-hour Robotaxi service is close. Replying on X to a rider who wanted Cybercab trips all night, he wrote that the capability would arrive “next month or so” once “the next tech to merge on the v15 plan” is ready.
The comment landed on September 4, one day after Tesla opened public Cybercab rides in Austin. It is the clearest near-term timeline yet for overnight unsupervised operation. Tesla’s paid Robotaxi network currently runs from 6 a.m. to 10 p.m. seven days a week across Austin, Dallas, Houston, Miami, Orlando, and Tampa.
next month or so. the next tech to merge on the v15 plan will enable it.
— Ashok Elluswamy (@aelluswamy) September 4, 2026
That 16-hour window is shorter than the 6 a.m. to 2 a.m. schedule the company used for much of the prior year.
Elluswamy did not name the specific feature or say whether the change would apply first to purpose-built Cybercabs, the existing Model Y fleet, or both. He also offered no city-by-city rollout list. The link to Full Self-Driving v15 is nevertheless significant.
Tesla has described v15 as a step-change architecture with seven parallel improvement tracks and roughly ten times more parameters than earlier builds. Early versions of that software already operate on the Robotaxi fleet and contain about 40 percent of the planned gains.
By July 2026, the unsupervised fleet had logged more than 380,000 miles across six cities in two states with what the company called an impeccable safety record and no notable incidents caused by the vehicles themselves. Tesla has repeatedly argued that camera-based end-to-end neural networks, rather than extra sensors, are the core of the solution.
Overnight service would test that claim in lower-light conditions and would also raise vehicle utilization, a key variable for Robotaxi unit economics. The company has already begun using public Superchargers at night and is building dedicated Robotaxi charging sites.
Riders have asked why software must change if the cars already drive in the dark. The practical answer appears to be reliability and scale: Tesla has held back mass expansion until more of the v15 stack is merged, citing the need for higher confidence before putting thousands of unoccupied vehicles on streets around the clock.
If the next module arrives on the timetable Elluswamy sketched, 24-hour service could begin in October 2026 in at least some markets.
That would mark a shift from a daytime-bounded pilot to a service that can run whenever demand exists, including the late-night hours that have so far remained out of reach.
Tesla is beginning to roll out Full Self-Driving Supervised v14.3.9 with a new active safety layer that can take control even when the driver is operating the car manually.
Tesla AI said the software can activate FSD on the driver’s behalf when an imminent collision is detected and Automatic Emergency Braking may not be enough. It may also engage if the system detects heavy distraction or an accidental FSD disengagement.
FSD Supervised v14.3.9 starting to roll out shortly
This release includes a new active safety feature set: FSD Supervised can now activate on your behalf when an imminent collision is detected and Automatic Emergency Braking (AEB) may not be enough.
It may also engage if we…
— Tesla AI (@Tesla_AI) September 4, 2026
The capability is essentially Automatic Collision Evasion. However, unlike conventional AEB, which mainly applies the brakes in a straight line, this feature can use steering, braking, and acceleration together if the car calculates that stopping alone will not prevent impact and a safer path exists. The system may change lanes or move toward a shoulder when conditions allow, then continue driving after the immediate threat is handled rather than simply coming to a stop.
The intervention is meant as a last-resort safety net, not a replacement for attentive driving.
Tesla Full Self-Driving v14.3.7 early review: FSD saved me from an accident
Tesla’s own description still frames FSD as supervised assistance. Secondary reports on internal release notes say the feature can fire while the car is being driven manually if cabin-camera monitoring suggests the driver is not sufficiently attentive, such as reaching toward the back seat, or if FSD appears to have been turned off unintentionally.
After the emergency maneuver, the car is expected to alert the driver and request a return to manual control.
The safety case is straightforward. Many collisions happen in the last second because a driver is looking away, fumbles a control, or faces an obstacle that braking cannot fully solve. A system that can both recognize that AEB is insufficient and execute a coordinated evasive path can reduce those remaining high-severity events.
Re-engaging after accidental disengagement also addresses a practical failure mode: a small steering nudge that drops FSD at the worst moment. The advantage is a background safety net that uses the same vision stack already running in v14, instead of leaving the car solely to emergency braking once the driver is no longer in command.
The feature still depends on FSD being enabled and, according to reports, an active FSD purchase or subscription. It does not make the vehicle unsupervised. Drivers remain responsible, and Tesla has not published how often the system is expected to intervene or how it will handle false positives.
If the rollout is conservative and the false-alarm rate stays low, the update is a meaningful step: FSD is no longer only a feature the driver turns on. In the rare moments when disaster is already forming, it can step in.
Tesla crosses major Unsupervised Self-Driving milestone
Tesla Robotaxi will be a 24/7 service: here’s when
