Comma.ai Under Federal Investigation After Crashes Linked to Stopped Vehicles
NHTSA's Office of Defects Investigation has opened an investigation into Comma.ai after five reported crashes involving its aftermarket hands-off driver assistance technology. Two of those crashes caused three deaths, according to the agency's account described by TechCrunch.
The crashes under review involved Comma devices failing to detect or respond to slow or stopped vehicles in the same lane. This is a known hard case for combined camera-and-radar systems, especially for forward speed control, where the system must separate a stopped car from roadside signs, road furniture and overhead structures like bridges.
As many as 11 people were injured across four of the crashes, some seriously. That leaves one of the five crashes with no reported injury, a distribution that points to high-energy impacts rather than low-speed contact.
A complication involves software provenance. Some crashes, including at least one fatal crash, may have involved forked or modified versions of Comma.ai's software, according to the ODI. Comma.ai's software is called openpilot, and it can be freely copied and modified. Third-party forks can change braking tuning, detection thresholds, when the system engages, and how strictly it enforces driver monitoring.
The February 2026 crash in Ascension Parish, Louisiana, involved a 2022 Toyota RAV4 running FrogPilot that struck a stopped first responder vehicle. Two rear-seat passengers in the Toyota died, in NHTSA's description. FrogPilot is a third-party fork of openpilot, not the standard build distributed by Comma.ai.
Comma.ai devices and openpilot are an aftermarket retrofit for two functions: Adaptive Cruise Control, which controls speed and following distance, and Automated Lane Centering, which controls steering to stay in lane. On supported vehicles, once installed and engaged, they provide sustained steering and speed assistance. Comma.ai warns drivers they must always be able to retake manual control immediately by stepping on either pedal or pressing the cancel button. Its device also uses a camera-based driver monitoring system to watch for distracted or sleepy drivers.
Comma.ai was founded in 2015 by George Hotz. Hotz stepped back from day-to-day operations in 2022.
The broader context here is liability and control in an open, aftermarket model. Factory-installed assistance ties sensors, controls, testing and software updates to a single vehicle and a single accountable manufacturer. Comma's approach separates the assistance system from the vehicle. Owners install third-party hardware, run open-source software, and can replace it with community-maintained variants. For engineers used to locked bootloaders, signed builds and traceable software bills of materials, which list every software component, the lack of a clear record of what software was running is the central technical issue.
In my view, the investigation will likely turn on three questions. First, whether stock perception and planning had a systematic weakness around stopped in-lane vehicles. Second, whether driver monitoring and takeover warnings were sufficient and enforced during the relevant drives. Third, where responsibility sits when a fork is involved. That third question is the most consequential for the wider ecosystem, because open source works well for infrastructure software where failures stay contained and can be patched, while on public roads a fork that weakens attention checks or changes braking behavior creates immediate physical risk even if the upstream project is conservative.
Worth flagging alongside that concern is what the model enables when it works. Low-cost retrofits bring lane centering and adaptive cruise to cars that shipped without them, and openpilot offers a level of transparency into disengagements, model behavior and driver state that closed systems rarely match. The path forward is probably more inspection, not less: stronger proof of which build is running, clearer in-cabin signals about which software variant is engaged, and monitoring that cannot be easily muted by a downstream build.
