Table of Contents
Hands-Off Driving Investigation: Two Fatal Crashes
A federal hands-off driving investigation into Comma's openpilot devices follows five crashes, two fatal. What parents of teen drivers should take from it.
On September 23, 2026, NHTSA’s Office of Defects Investigation opened case PE26007 into Comma.ai’s driver-assistance devices after five reported crashes. The agency’s finding was specific: the devices “fail[ed] to detect or respond to slow or stopped vehicles in the same road lane.”
In one of those crashes, in February 2026 in Ascension Parish, Louisiana, a 2022 Toyota RAV4 struck a stopped Gonzales Police Department vehicle in the left lane with its emergency lights on. Two rear-seat passengers died. The vehicle was running FrogPilot, a third-party modified version of Comma’s software.
That last sentence is why this hands-off driving investigation matters beyond one company.
Key Takeaways
- NHTSA’s Office of Defects Investigation opened PE26007 after five reported crashes involving Comma’s aftermarket driver-assistance devices.
- The identified failure mode is the oldest known weakness in adaptive cruise control: not reacting to slow or stopped vehicles already in the lane.
- The fatal Louisiana crash involved FrogPilot, a community-modified fork of openpilot, not the manufacturer’s own release, an accountability gap with no clean regulatory answer.
- Comma’s devices are aftermarket: a box you buy and install, controlling the car’s existing adaptive cruise and lane-centering functions.
- IIHS research (Cicchino, 2025) “did not find any crash-reduction advantage for vehicles equipped with partial driving automation” over comparable models with ordinary crash-avoidance features.
What Comma actually sells
Comma.ai, founded in 2015 by George Hotz, makes aftermarket devices that control a car’s existing Adaptive Cruise Control and Automated Lane Centering systems. Its software, called openpilot, can “steer, accelerate, and brake automatically for other vehicles within its lane.” TechCrunch’s reporting describes it as functioning similarly to Tesla Autopilot and GM Super Cruise. The device includes a camera-based driver-monitoring system to watch for distracted or sleepy drivers.
Two words in there deserve emphasis. Aftermarket means the hardware was not installed or validated by the automaker; it plugs into a car’s existing network and issues commands. Within its lane means this is lane-keeping and following, not autonomous driving.
The phrase “hands-off driving” in the headlines is loose. A system with camera-based driver monitoring that watches for inattention is by construction a hands-on, eyes-on system. In SAE terms, as IIHS summarises them, this is Level 2, where “the human remains responsible” and the system “cannot ever replace the driver.” The marketing language and the engineering reality are not aligned anywhere in this product category, which is a large part of why crashes like this keep happening.
Why stopped vehicles are the hard case
Adaptive cruise control usually relies on radar to measure the distance and closing speed of the vehicle ahead. Radar is excellent at that job and has one structural problem: the world is full of stationary metal. Bridges, overhead signs, guardrails, parked cars, manhole covers. If the system braked for every stationary radar return, it would be unusable.
So these systems filter. Classically, they weight moving targets heavily and discount stationary returns, which works beautifully right up to the moment the thing ahead of you is a car that has stopped. The radar sees it as clutter. The system does not brake.
This is not an obscure edge case or a novel defect. It has been the documented weakness of radar-based ACC since the technology shipped, which is why camera-based detection and sensor fusion exist, and why automatic emergency braking is tested separately from cruise control. NHTSA lists Automatic Emergency Braking and Forward Collision Warning as distinct technologies from Adaptive Cruise Control in its driver-assistance guidance for exactly this reason: they solve different problems.
A stopped police vehicle with emergency lights activated is the worst version of the case, and it recurs across this entire product category. The lights add visual noise. The vehicle is stationary. The scene is often partly in a travel lane.
The fork problem nobody has solved
Here is the part that has no clean answer. openpilot is open-source software. FrogPilot is a third-party modified version of it. The vehicle in the fatal Louisiana crash was running FrogPilot.
So trace the chain of responsibility. The automaker built the car and its ACC and lane-centering. Comma built a device and software that command those systems. An unaffiliated developer modified that software. A driver installed the modification. A federal agency is investigating the company whose code was forked.
Every link in that chain is doing something defensible in isolation. Open source has produced enormous value in software, including in safety-adjacent fields. Aftermarket automotive parts are a legitimate, regulated industry. But the combination, community-modifiable code issuing steering and braking commands to a two-tonne vehicle on a public road, sits in a regulatory space that was designed for neither open-source software nor aftermarket electronics.
Comma.ai “could not immediately be reached for comment,” and Hotz “did not immediately respond to requests for comment,” per TechCrunch. The investigation is a preliminary evaluation, not a finding. An opened docket is not a defect determination.
Three products that look the same and are not
| Factory ADAS (e.g. GM Super Cruise) | Aftermarket device (Comma openpilot) | Community fork (e.g. FrogPilot) | |
|---|---|---|---|
| Who validated it | The automaker, for that model | The device maker, across many models | Nobody with liability |
| Covered by vehicle warranty | Yes | No, and may void parts of it | No |
| Subject to recall process | Yes | Equipment recalls possible | No mechanism |
| Who pushes updates | Automaker | Device maker | The fork’s maintainer |
| Driver monitoring | Varies; some camera-based | Camera-based | Can be modified |
| SAE level | 2 | 2 | 2 |
| Who is responsible in a crash | Driver, with manufacturer exposure | Driver, with unclear device exposure | Driver, essentially alone |
The bottom row is the one to read twice. In all three columns the driver is responsible. What changes is how much institutional accountability sits behind the software, and in the third column, the answer is close to none.
What the research says about partial automation generally
This is where the honest picture gets uncomfortable for everyone, including the automakers.
IIHS states plainly: “While partial driving automation may be convenient, we don’t know if it improves safety.” Its research with the Highway Loss Data Institute “yielded mixed findings about any safety benefits beyond the crash avoidance features typically included.” More pointedly, researchers “did not find any crash-reduction advantage for vehicles equipped with partial driving automation” compared with similar models that had only crash-avoidance technology (Cicchino, 2025).
On driver monitoring, IIHS notes that automakers “use various driver monitoring and alert strategies to help drivers return their attention to the road” (Mueller et al., 2025a) but that “some of these strategies work better than others” (Mueller et al., 2021). And drivers may become “disengaged because the vehicle is handling more of the driving” (Reagan et al., 2021).
The institute also records that “system misuse has already been implicated in fatal crashes,” citing NTSB investigations from 2017, 2019, 2020 and 2026.
Put together: the crash-avoidance features, automatic emergency braking, forward collision warning, have measurable benefits. The layer on top that steers and follows has not demonstrated an additional crash reduction, and introduces a disengagement problem. That is not an argument against the technology’s future. It is an accurate description of where the evidence sits in 2026.
What to do at home
Find out exactly what your car has, by name
Open the owner’s manual and write down the actual feature names. NHTSA’s guidance is blunt about this: review the manual “for more information on your vehicle’s technology and safety features.” Forward collision warning only warns, it “does not take action to avoid a crash.” Automatic emergency braking acts. Adaptive cruise control follows. Lane keeping assist nudges. Those are four different things and most drivers conflate them.
Teach the stopped-vehicle rule explicitly
If your teen drives a car with adaptive cruise control, tell them the specific limitation: these systems are weakest at reacting to something already stopped in the lane. Construction, a crash scene, a stalled car, a police vehicle. Cruise control off in those conditions. That is one sentence and it addresses the exact failure mode in PE26007.
Say no to aftermarket driving automation on a teen’s car
This is the one place in this article where I would be prescriptive. A community-modified driving stack on a car driven by an 18-year-old combines the least validated software with the least experienced driver. There is no upside that justifies it.
Watch for the quiet substitution
The risk with partial automation is not a dramatic failure. It is that a driver gradually stops scanning because the car has handled it a hundred times. Reagan et al. (2021) documented that disengagement. With a new driver, that habit forms before any real judgement does. We wrote more on this in learning to drive with driver assistance switched on.
What not to do
Do not read “federal investigation” as “proven defect.” PE26007 is a preliminary evaluation. NHTSA opens these to gather information, and many close without a recall. The facts worth acting on are the ones already established: the failure mode is real and documented, the fatal crash involved modified software, and the driver was responsible for the vehicle the entire time.
What to Watch For Over the Next 3 Months
- Week 4: Watch whether PE26007 escalates to an Engineering Analysis. PE is the first stage; EA means the agency found enough to look harder. That transition, not the opening, is the meaningful signal.
- Month 2 red flags: Watch for any statement from Comma about whether forked versions of openpilot are supported, detectable, or disclaimed. The accountability question for community modifications is the genuinely new issue here, and whoever addresses it first sets the template.
- Month 3 self-check: Ask your teen driver, cold, what their car’s cruise control will not do. If they cannot name the stopped-vehicle limitation, the feature is operating with more trust than it has earned.
Frequently Asked Questions
Is this about Tesla?
No, though the article notes that “Tesla and Ford have faced similar investigations after their systems were involved in multiple crashes.” PE26007 concerns Comma.ai’s aftermarket devices running openpilot. The failure mode is shared across the category rather than unique to any brand.
What does “hands-off driving” actually mean here?
Loosely, and that is part of the problem. Comma’s device includes camera-based driver monitoring that watches for distraction or drowsiness, which means it expects an attentive driver. In SAE terms it is Level 2, where the human remains responsible and the system cannot replace the driver. The breakdown of the levels is in the driving automation levels explained for parents.
Can anyone install one of these on any car?
Comma’s devices work with a range of vehicles that already have adaptive cruise control and lane centering, controlling those existing systems. They are aftermarket, meaning they are not installed or validated by the automaker and are not covered by the vehicle’s warranty or recall process in the way factory equipment is.
Why didn’t the car brake for a police vehicle with its lights on?
The agency’s finding is that the devices failed “to detect or respond to slow or stopped vehicles in the same road lane.” Radar-based following systems routinely filter out stationary returns to avoid braking for signs and bridges, which makes an already-stopped vehicle the hardest case in the entire problem space.
Should I turn off my own adaptive cruise control?
Not necessarily. Use it where it was designed to work: flowing traffic on limited-access roads. Switch it off approaching anything stopped or irregular, construction, incident scenes, heavy congestion that has come to a halt. That is where the evidence says the weakness lies.
Does this mean driver assistance is useless?
No. The crash-avoidance subset, automatic emergency braking and forward collision warning, has demonstrated benefits and NHTSA recommends these technologies. The finding from Cicchino (2025) is narrower and more specific: adding the steering-and-following layer on top did not show an additional crash reduction in the data studied.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- TechCrunch. (2026, September 23). “Comma’s hands-off driving tech under investigation after 2 fatal crashes.” https://techcrunch.com/2026/09/23/commas-hands-off-driving-tech-under-investigation-after-2-fatal-crashes/
- Insurance Institute for Highway Safety. “Advanced driver assistance” (citing Cicchino, 2025; Mueller et al., 2021, 2025a; Reagan et al., 2021; NTSB, 2017–2026). https://www.iihs.org/topics/advanced-driver-assistance
- National Highway Traffic Safety Administration. “Driver Assistance Technologies.” https://www.nhtsa.gov/vehicle-safety/driver-assistance-technologies
- National Highway Traffic Safety Administration. “Recalls.” https://www.nhtsa.gov/recalls
- Insurance Institute for Highway Safety. (2026, July 23). “How safe are driverless cars?” https://www.iihs.org/news/detail/waymos-driverless-cars-crash-less-often-than-people
- Wikipedia contributors. (2026). “Self-driving car.” https://en.wikipedia.org/wiki/Self-driving_car
