When you think about it, it’s kind of cool that computers can control a mainstream vehicle’s steering, power, and brakes in a vague facsimile of riding alone with a learner driver. It’s definitely not the full Knight Rider treatment, but speed control that follows the car in front of you, and hands-off lane-keeping assistance is nice to have on long, tiring highway slogs. Level 2 autonomy has its flaws, however, and a new NHTSA investigation into open-source ADAS provider Comma.ai is worth watching.
The principle behind the Comma.ai device is rather interesting: It’s a unit that takes over the steering, braking, and powertrain systems that are already in your modern car and promises hand-free semi-autonomous driving. Classified as a Level 2 system, it requires constant monitoring and accepts a variety of open-source driver assistance systems. Openpilot is the most popular, but several other profiles have sprung up over the years.
As this is an open-source system that’s still in constant development, some drivers who’ve tested Comma.ai and Openpilot over the years have found some limitations in the capabilities of its integrated sensor suite. In a 2026 test, tech YouTuber Linus Tech Tips found it would often travel well below the posted speed limit. In a much earlier 2020 Car and Driver test, the magazine noted lows including “A too-large and unadjustable gap from cars ahead, slows substantially for curves, flashes unnecessary warnings.” It’s not all bad news, though; In 2020, Consumer Reports tested 17 manufacturer-installed advanced driver assistance systems and version two of the Comma.ai unit (the latest model is version four), and this little aftermarket device running Openpilot came out on top. It excelled in performance, driver engagement, and ease of use, making it a viable option for owners of older cars with primitive ADAS systems.
Unfortunately, “viable” doesn’t mean “perfect.” On Sept. 21, the National Highway Traffic Safety Administration opened an investigation into the three latest versions of the Comma.ai device after a series of serious collisions involving the use of the system.
The Office of Defects Investigation (“ODI”) has identified crashes involving vehicles equipped with devices manufactured by comma.ai (“comma”) occurring under similar circumstances. ODI received notice of five (5) crashes in which vehicles with comma devices installed reportedly struck stopped or slowly moving vehicles within the same travel lane. ODI has confirmed that the systems were powered on or engaged in multiple crashes. Preliminary data reviewed by ODI suggests that the comma system may not have adequately detected or responded to the in-lane vehicles. Two crashes resulted in a total of three (3) fatalities. Four of the five crashes contributed to various non-fatal injuries, including serious injuries, to eleven (11) other individuals.
This seems to exhibit a similar failure pattern to several other advanced driver assistance systems that are or were under federal investigation. In 2025, NHTSA opened an investigation into Ford vehicles with Blue Cruise hands-free highway driving assistance after two fatal collisions occurred in which Mustang Mach-E electric crossovers operating on Blue Cruise struck stopped vehicles on controlled-access highways. As per NHTSA:
In both fatal collisions referenced in the PE24012 opening resume, the subject Ford Mustang Mach-E vehicle was traveling over 70 mph on a controlled-access highway during nighttime lighting conditions with hands-free BlueCruise engaged when it collided with a stationary vehicle. Analysis of data imaged from the vehicles’ event data recorders demonstrates that in each incident, the driver did not apply the brakes or take evasive steering action, and no deceleration was initiated by either the BlueCruise system or PCA prior to impact. Through the agency’s crash analysis, four additional frontal collisions were identified where the subject Ford impacted a stopped or slow-moving lead vehicle or another stationary object located in the travel lane.
Earlier still, NHTSA opened an investigation into Tesla vehicles equipped with the Autopilot advanced driver assistance system for a similar but slightly different reason. According to a 2022 document, “The investigation opening was motivated by an accumulation of crashes in which Tesla vehicles, operating with Autopilot engaged, struck stationary in-road or roadside first responder vehicles tending to pre-existing collision scenes.”
As it stands, the Ford investigation is still ongoing, while the Tesla investigation escalated to an engineering analysis before being formally closed by a huge recall of more than two million vehicles. Interestingly, the Tesla recall mainly pertained to flaws in the driver monitoring system meant to keep track of whether drivers were actually paying attention. As the recall report states, “In certain circumstances when Autosteer is engaged, the prominence and scope of the feature’s controls may not be sufficient to prevent driver misuse of the SAE Level 2 advanced driver-assistance feature.”
While a conclusion to the Comma.ai investigation is likely some ways out, a possible future recall might get weird. The difference here is that Comma isn’t an automaker, it’s a company that makes an aftermarket device that runs open-source software, and what this means for potential enforcement if NHTSA concludes that a fix is necessary is yet to be seen.
See, open source software presents a unique challenge in the event of a recall, because it tends to foster a fair and legal ecosystem of forks and spin-offs and alternate versions. If the issue lies within the software version itself, it may be possible to update standard Openpilot, but certain other forks already in development may or may not follow suit on a similar timeline. Experimental features may continue to exist, and Comma.ai doesn’t have full control over what other developers do with its open-source code.
Of course, there’s also the possibility that the NHTSA investigation finds Comma.ai’s driver monitoring system to be sufficient, and the problem to exist between the seat and the steering wheel. An inherent limitation with Level 2 ADAS is that these systems require constant monitoring, and if a driver fails to pay attention and then fails to heed the system’s warnings, the success of safely bringing things to a stop is limited across all ADAS systems. Stopping in a lane isn’t always an ideal solution because it can result in a rear-end collision; pulling over isn’t really an option if there’s no hard shoulder; and disabling driver assistance functions until the next ignition cycle probably isn’t ideal if the driver is incapacitated.
In some ways, this investigation into a small, independent ADAS company might be Level 2 autonomy’s biggest test yet because it’s not a closed, proprietary ecosystem. Granted, Comma.ai has survived investigation before. In 2016, the company decoupled Openpilot from its Comma hardware after a NHTSA inquiry, requiring users to load it onto the device after acquiring said unit. Let’s see what happens this time.
Top graphic image: YouTube/Comma
It’s almost like systems that require driver attentiveness but do everything possible to disconnect the driver from the process of actually driving are a bad idea. Who could possibly have guessed that?
Genie is out of the bottle here. The faster L3 is implemented, the sooner L2 can be outlawed.
Best result would be the operator is ruled 100% responsible for damages, OEM system or aftermarket, and the gov does an ad campaign saying as such.
I don’t know what hardware they were running but there has been some posts on twitter recently with people experimenting with having general AI (like Chat GPT) drive a toyota around a parking lot. Apparently, they had to prompt around the AI as it didn’t want to do it (apparently adding the word sand box to the prompt did the trick for at least one) , but they got several to work, feeding the AI a live camera feed and telling it to move to a certain location.
Why? Who “needs” this? I’ve driven for years in metropolitan areas with my FEMA laptop, two cell phones and my lunch with better control and situational awareness. Putting your life in the hands of an open source, non-Lidar system that sidesteps all manufacturer liability is just dumb, more so considering that most of us reading The Autopian likely are above the average skill set for American drivers.
I’d feel a lot better about so-called self-driving cars if there was some external indicator to show that FSD is turned on. A red warning light that other drivers can see, or something along those lines.
They’ve already picked on turquoise as the autonomous warning lamp colour, instead of the much more sensible purple.
I was going to suggest blue actually, but then I remembered that in most places it’s illegal to have blue lights on your car.
I can appreciate the work and the idea, but I want this far away from my cars. Also, open source is great in theory and in lots of implementations, but eventually someone will hack this and before you know it your car is at Taco Bell. That happens now, but usually other bad decisions happened before that.
There’s safety items hard coded in to the Comma that can’t be changed by the open source software. It also doesn’t have an active internet connection most of the time (mine is linked to my home wifi only for updates and doesn’t have internet access while driving) and is easy to disengage the system, so I don’t see it being hacked. It’s also very easy to take over the steering from the Comma (it’s not very forceful and doesn’t fight your input or override).
Which car do you have the system on?
I’ve only ever talked to one other person that had a Comma.ai, and it was on a recent Lexus ES 300h.
I have it on a 2021 Ram 1500. I’m running Sunnypilot on it.
The thing about open source is there are usually far fewer of these vulnerabilities than closed source, and when said vulnerabilities are found they are patched far faster.
This is only internet-connected for updates. It is not possible to “hack” the device, and it can’t engage itself. There are limitations hard-coded into the hardware which cannot be overridden.
None of the vulnerabilities you mentioned are actually possible on a Comma.
This is why OEMs lock up their software, and have enormous validation processes. Making every application of this technology potentially a prototype install makes assigning liability very difficult and validation impossible.
Still, it’s exciting to be a control subject in someone else’s experiment.
So I have two Comma 4s myself – one in my Ioniq 5 and one in my Kona electric, both running Sunnypilot. My partner has one in his Ioniq 6 running Starpilot.
This was a good article, but it’s worth mentioning that the Comma doesn’t always take over both longitudinal (distance) and lateral (steering) control. The two systems can be used together, or independently, so the car can steer while the driver controls the speed, or the car can control the speed while the driver steers.
Also, not all vehicles support both longitudinal and lateral control. None of our Hyundais support longitudinal control, so that’s still done with the car’s stock system. I am not sure which vehicles were involved in these accidents, and it’s mentioned that Frog Pilot was used in one of them. It’s also not even clear whether the Comma was engaged at the time of the accidents.
Unlike Tesla’s Full Self-Driving (Supervised), which actively encourages driver inattention through deceptive marketing and the manner in which the system works, Comma makes no such claims. It is clearly marketed as a driving assistance system which requires driver attention at all time. There is a camera pointed directly at the driver which will start yelling at the driver if they aren’t paying attention, and there is no way to defeat that, not even with other branches.
Most of us in the Comma community basically only use it on highways, where it really shines. Yeah, you’ll see YouTubers like Linus stretching the capabilities of the device and pointing out how it doesn’t work well in complex situations with dense traffic. That just isn’t the best way to use a Comma (although it makes more revenue on YouTube compared to a boring video of a Comma-equipped vehicle driving for hours handsfree on the highway).
These accidents were 100% driver error. I don’t need an investigation to tell me that. As mentioned in the article, ADAS systems cannot always detect when the vehicle might need to suddenly swerve and/or stop to avoid an accident. That is the human driver’s job, and Comma never obfuscates that fact. I suspect that the drivers might be trying to shift liability away from themselves, especially since one of the accidents seems to have involved the tragic loss of human life.
I absolutely love my Commas. They make driving on highways extremely relaxing. The lane centering and automatic lane changes are pretty much flawless and very human-like, plus unlike most factory ADAS systems, it handles construction zones and driving in bad weather really well.
I certainly hope this investigation assigns blame where it belongs. Commas are a way to access amazing ADAS on a device you own 100% with open-source software that can be taken from one vehicle to another and doesn’t require a subscription. It’s much cheaper than Blue Cruise or Super Cruise while being far superior to both. I feel bad for Comma.ai, and really hope the investigators take their time and do justice for the people killed and injured by blaming the drivers, not this device.
You said everything I wanted to say better than I could have. I have a Comma 4 running on my Ram 1500 and it’s great. I’ve towed my Airstream almost 4000 miles with it this year, and probably have driven another 4,000 without a trailer. It’s not perfect, but also, I’m actively monitoring and have learned when it is not confident in the decisions. I feel it does a better job for me than Blue Cruise or Super Cruise (I’ve driven both). The Ram Adaptive Cruise Control is being used by the Comma as well to keep distance from vehicles and adjust speed. If the vehicles have forward collision avoidance systems, those should come into play.
I treat these systems kind of like being a passenger with a teen driver who is still learning. They do a pretty good job most of the time, but you have to be a second set of eyes, especially in tricky situations like residential neighborhoods where a kid or dog could run out from behind a vehicle, or in odd road arrangements. I’m always ready to take over. I find I can actually pay a lot more attention to what is going on around me, because I’m not concentrating on keeping the vehicle centered in the lane. I watch my mirrors a lot more than I would without Comma driving. Comma also does have driver monitoring and it does yell at you if you aren’t paying attention to the road. You don’t have to hold the steering wheel, but you do need to pay attention.
I love my Comma 4. I don’t think I could have pulled my camper from Michigan to North Dakota and back and not been fatigued without it. I love the engagement of driving and have a manual 2006 Miata that I drive most of the summer, but I’m happy to let a device take over keeping my truck centered in a lane on 1000 miles of interstate driving at 70 mph.
I also find that I tend to be more situationally aware when driving with my Comma engaged. And I definitely am less fatigued when on long journeys.
It’s nice to hear that it does well with towing. It must have been a lifesaver on such a long journey!
This is fair. I have no experience with Comma.ai, although theoretically I could have on a few cars I’ve purchased over the years.
I had Super Cruise on a Cadillac Lyriq and it was decent enough; I didn’t have any issues with it or its limitation to mapped, divided-access highways. I also liked that, thanks to GM’s Safety Alert Seat that debuted years earlier, you could set Super Cruise to vibrate the seat for system alerts, instead of having it binging and bonging at you. And it was far better than the versions of Ford’s BlueCruise I tested on the Mustang Mach E and the F-150.
Yesterday, I took delivery of a new Rivian R2, which (being the Launch Edition) is bundled with Autonomy+ for life. Unlike Super Cruise, it doesn’t rely upon maps and can theoretically be used anywhere it can read the lane markings. It can also already read street lights and traffic signs, and is showing those in the instrument cluster. Rivian is targeting a late-2026 or early-2027 customer rollout of full point-to-point autonomy, including navigating stops and intersections…but I personally plan to just use it on the highway and interstate.
The Comma doesn’t require roads to be mapped, so it’s more similar to the Rivian system. I’d be interested to see how they compare. In my experience. both Super Cruise and Blue Cruise are a far inferior experience compared to my Comma. They aren’t even close.
I think the real danger here is any kind of ruling that stops self-repair. I worry that this will be the first one brought to the courts because it’s not backed by a manufacturer with tons of money and connections. But if they say you can’t do this, it’s not a leap to think the ruling will be owners can’t touch these systems at all. Maybe I’m just being paranoid…
Anyway… I kinda wish geohot had stuck to no money firmware cracks. The world was more fun when that was his interest.
This will be more complex due to the different implementations among different vehicles. For example, in some vehicles the Comma can take over fairly completely from the factory ADAS. Adaptive cruise and lane keeping handled by the Comma. In others, the Comma acts in concert with the factory ADAS where it handles only a part such as lane keeping while the vehicle handles adaptive cruise. Some forks on some vehicles even offer modes where the driver handles speed while the Comma handles lane keeping. Driver monitoring is aggressive. More than a few seconds of looking away or if it detects a cell phone in hand and it makes a big racket.
Further muddying things are the different forks off of Openpilot (Sunnypilot, BluePilot, etc.) and the different builds within each fork. Some people are wanting to test features while others stick with a stable release.
I purchased a Comma 4 earlier this year. It works very well on long straight roads like interstates. That said, Comma makes it very clear that this device isn’t a replacement for being attentive.
An excellent comment. There’s so much information we don’t know about these accidents. As you mentioned, it’s basically impossible to be distracted while using the Comma. The driver monitoring is really good.
What kind of vehicle do you use it in?
I looked into comma.ai a year or so ago for my expedition max but after discussing with the developers found that they were not going to support the earlier expeditions in that generation.
Having driven around with autopilot for several years, I do think, as long as these systems are being treated as cruise control +, then they aren’t inherently unsafe. As long as the system fails safes to off and the driver is ready to take over
How well does the safety record of this system compare to the same number of miles driven by human drivers?
Great question. I’m not sure there’s much data, especially since the Comma isn’t necessarily “on” or “off” like most ADAS systems. MDAS offers blended modes. It would be a difficult question to answer.
I can’t imagine the answer would be anything other than finding a better safety record when the Comma is engaged. It would just be difficult to get that data, especially since it’s not collected in a single location by a vehicle manufacturer. All of that data lives on the device and can only get exported if the user manually pulls it.
Right! We, the humans most likely cause more damage while drinking mocha lattes and texting while driving cars without level 2 help. Bring back the manual to give people something to do with there damn hands 🙂
A quote in this article says V2 of ai, was voted the best due to performance, driver engagement, and ease of use.
Isn’t the whole point of autonomous driving vehicles zero driver engagement? Or like golf did they win in the driver engagement category by having the least driver engagement?