# comma.ai > Source: https://aiwiki.ai/wiki/comma_ai > Updated: 2026-06-09 > Categories: AI Companies, Autonomous Vehicles > From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution. # comma.ai **comma.ai, Inc.** is an American advanced driver-assistance technology company headquartered in San Diego, California. Founded in September 2015 by security researcher and entrepreneur [George Hotz](/wiki/george_hotz), the company develops aftermarket hardware devices and the open-source openpilot driving software, which adds Level 2 longitudinal and lateral control to a growing list of supported production vehicles.[^1][^2] comma.ai positions itself as a crowd-sourced, open alternative to closed driver-assistance stacks from automakers and from systems such as Tesla Autopilot, distributing both its driving policy code and its hardware schematics under permissive terms.[^3][^4] As of 2025 openpilot is reported to upgrade the [advanced driver-assistance systems](/wiki/adas) of 300 or more supported car models, with users having accumulated more than 100 million miles of openpilot driving.[^3][^5] The company is structured as a small for-profit hardware business that sells a flagship device (currently the comma 3X, with the smaller comma four announced for late 2025) together with a wiring harness and mounting kit; the software running on the device is released free of charge under the MIT License.[^3][^4][^6] Beyond commercial product work, comma.ai has produced research artifacts including the 2016 paper "Learning a Driving Simulator," and its founder later spun out the [machine learning](/wiki/machine_learning) toolchain effort that became tinygrad, used both for openpilot inference and as a stand-alone framework.[^7][^8] | Attribute | Value | | --- | --- | | Founded | September 2015[^2] | | Founder | George Hotz (geohot) | | Headquarters | San Diego, California | | Industry | [Driver assistance](/wiki/autonomous_vehicle), automotive software | | Key product (hardware) | comma 3X (2023); comma four (announced 2025)[^6][^9] | | Key product (software) | openpilot (open source, [MIT License](/wiki/mit_license))[^3][^4] | | Funding | ~$8.1M disclosed across seed and Series A rounds[^10][^11] | | Lead investor | [Andreessen Horowitz](/wiki/andreessen_horowitz) (2016 seed)[^10] | ## History ### Pre-history: George Hotz before comma.ai The company is closely identified with the public profile of its founder. George Hotz, who goes by the alias "geohot," first attracted broad attention in August 2007 when, at age 17, he became the first person reported to remove the SIM lock on an Apple iPhone, trading an unlocked unit to the founder of a wireless repair company for a Nissan 350Z and three additional iPhones.[^12] He subsequently developed several iPhone jailbreaking tools, including the blackra1n exploit released in October 2009, before announcing in July 2010 that he was discontinuing iPhone work.[^12] In late 2009 and 2010 Hotz turned to Sony's PlayStation 3 platform. On January 22, 2010 he announced that he had achieved read and write access to the console's system memory and hypervisor-level access to its CPU, and on January 2, 2011 he published a copy of the PS3 root signing key on his personal website.[^12] Sony filed for a temporary restraining order on January 11, 2011, and the case was settled out of court in April 2011 under terms requiring Hotz to refrain from further work on Sony products.[^12] Hotz briefly attended both Rochester Institute of Technology and Carnegie Mellon University without completing a degree. Between the PS3 affair and comma.ai he worked at Facebook (May 2011 to January 2012), at Google's Project Zero security team starting July 16, 2014, and at the [machine learning](/wiki/machine_learning) startup Vicarious from January through July 2015.[^12] ### Founding and the Acura ILX demo (2015 to 2016) Hotz founded comma.ai in September 2015 to commercialize aftermarket driver-assistance hardware.[^2][^12] In December 2015 *Bloomberg Businessweek* published a profile of Hotz with video of a 2016 Acura ILX that had been retrofitted by his small team to drive on California's Interstate 280 under [neural network](/wiki/neural_network) control.[^2] The demonstration drew a cease-and-desist letter from the California Department of Motor Vehicles, which contended that comma.ai was testing an autonomous vehicle without the required permit.[^2][^12] In April 2016, [Andreessen Horowitz](/wiki/andreessen_horowitz) led a $3.1 million seed round at a $20 million pre-money valuation; partner Chris Dixon joined comma.ai's board.[^10] A 2018 funding filing later disclosed an additional roughly $5 million round, bringing total disclosed external funding to approximately $8.1 million; comma.ai's own communications repeatedly contrast this figure against the much larger sums raised by closed-source autonomous driving competitors.[^10][^11] ### The comma one, NHTSA, and the open-source pivot (2016) On September 13, 2016, comma.ai announced its first product, the comma one, at TechCrunch Disrupt; the device was a small windshield-mounted unit running a smartphone-class processor and was offered for a $999 hardware fee with a $24 monthly subscription, initially supporting select Honda and Acura models.[^2][^13] On October 27, 2016 the U.S. National Highway Traffic Safety Administration (NHTSA) issued comma.ai a "special order" letter, demanding written, sworn answers to 15 questions about the comma one's design, safety testing, and intended sale, on penalty of civil fines of up to $21,000 per day for non-response.[^13][^14] The next day, Hotz posted from the official comma.ai Twitter account that the company would cancel the comma one rather than continue down the regulatory path; *TechCrunch* and *CNN Business* both quoted Hotz characterizing dealings with regulators and lawyers as not worth the time.[^13][^14] On November 30, 2016 comma.ai released its full driving software stack as open source under the MIT License, branding the project openpilot and explicitly disclaiming any warranty.[^2][^3] The pivot effectively converted the commercial deployment risk into a community research project: by shipping the device only as a development kit and the software only as research-grade alpha code, comma.ai sidestepped the regulatory questions NHTSA had asked of a finished product.[^2][^13] ### EON and the dashcam form factor (2017 to 2018) On October 19, 2017, comma.ai began shipping the EON Dashcam DevKit, a $699 head unit built around a OnePlus 3T smartphone running a customized Android variant called "NEOS," paired with the comma.ai Panda (a USB-to-CAN interface) and a custom car harness.[^2] The EON exposed openpilot to a much wider population of cars than the comma one, since the harness and CAN messaging definitions could be ported by community developers; over the course of 2018 and 2019 supported vehicle counts grew to encompass most Honda, Toyota, Lexus, Hyundai, and Kia models equipped with stock lane-keeping or adaptive cruise systems.[^2][^15] A Series A funding round disclosed in April 2018 raised approximately $5 million and reportedly valued the company near $50 million.[^11] ### comma two (2020) The first batch of 300 comma two units became available on January 7, 2020 at a price of $999.[^16] The comma two preserved EON's smartphone-class compute (Snapdragon-based) while moving to a dashcam-style enclosure mounted directly to the windshield, drawing power through USB-C rather than from an internal battery and adding two infrared illuminators for nighttime driver-monitoring.[^2][^16] It also extended openpilot's hardware to a fourth CAN bus over the OBD-II connector and introduced a low-power "off" mode after three days of inactivity to avoid draining the host vehicle's battery.[^16] In November 2020 Consumer Reports ranked the comma two and openpilot above every other driver-assistance system it tested, including Tesla Autopilot, GM Super Cruise, and Ford Co-Pilot 360, citing driver-engagement and ease-of-use criteria; the comma two was the only aftermarket device in the test.[^17] ### comma three devkit (2021) and comma 3X (2023) The comma three was announced on July 31, 2021 and shipped as a development kit later that year. It moved from a phone-based architecture to a custom-engineered head unit built around the Qualcomm Snapdragon 845 system-on-chip (eight ARM cores, Adreno 630 GPU), a 6-inch OLED display at 2160 x 1080, three road-facing cameras with on-board image-signal processing, and a built-in panda safety microcontroller; storage options were 128 GB, 256 GB or 1 TB of NVMe SSD, with the 256 GB devkit priced at $1,999 and the 1 TB devkit at $2,499.[^2][^18] The comma 3X was announced at COMMA_CON 2023 and began shipping on October 12, 2023 at a base price of $1,250.[^2][^9] Externally similar to the comma three, the 3X was redesigned for manufacturability: the in-board panda microcontroller was upgraded to an STM32H7 (matching the standalone red panda), giving the 3X native CAN FD support so that it could communicate with the newer 2020+ vehicle electrical architectures used by manufacturers such as Hyundai-Kia, Ford, and General Motors.[^9] The road cameras were upgraded to the OmniVision OX03C10 sensor with 140 dB high-dynamic-range capability (up from 120 dB on the comma three), and comma.ai reported that the simplified 3X assembly halved the parts count and was designed to drop the production defect rate below the comma three's 3.9% baseline.[^9] The 3X also added a stereo microphone and speakers and used built-in Wi-Fi plus LTE for over-the-air software delivery.[^9] Compatible brands at launch covered Audi, Chevrolet, Ford, GM, Honda, Hyundai, Jeep, Kia, Lexus, Mazda, Subaru, Toyota, and Volkswagen Group vehicles.[^9] ### Leadership transition and the tiny corp spinoff (2022 onward) Hotz announced on October 19, 2022 that he was stepping down from day-to-day leadership of comma.ai, while remaining on the company's board; he later cited a desire to focus on his next venture and a sense that the company had matured beyond the stage at which his particular skills added the most value.[^2][^12][^19] On November 5, 2022 he formally founded the tiny corp, a small Delaware company whose first effort was the tinygrad [deep learning](/wiki/deep_learning) framework and whose intended business was selling AMD-GPU-based machine-learning compute appliances.[^8][^12] Between November 21 and December 21 of 2022 Hotz briefly worked as a 12-week "intern" at Twitter under newly installed owner Elon Musk, intending to repair the platform's search and login-prompt user experience; he resigned after roughly four to five weeks, posting that he did not believe he could have meaningful impact at the company.[^20] In May 2023 tiny corp announced a $5.1 million funding round to expand work on tinygrad and a related hardware product line.[^8] tinygrad and openpilot share authorship and infrastructure but are technically separate projects; tinygrad supplies the runtime in which openpilot's driving models execute on the Snapdragon GPUs of the comma 3X.[^8] ### comma four (2025) At COMMA_CON 2025, comma.ai announced the comma four, a successor that retains the comma 3X's compute and triple-camera sensor suite while shrinking the total enclosure by roughly a factor of five and placing the active display behind a 1.9-inch, 300 PPI OLED touchscreen.[^21] The comma four is priced at $999 (or $699 with a trade-in of an older comma device), runs the same Snapdragon 845-class processor branded as "Snapdragon 845 MAX" by comma.ai, and continues to be assembled in San Diego.[^21] ## Mission and business model In its self-presentation, comma.ai's stated mission is to "ship a self-driving car for $1000" by upgrading existing automobiles with low-cost aftermarket hardware and crowd-sourced software, rather than producing vehicles from scratch or pursuing a robotaxi service.[^2][^4] The business model is intentionally narrow: comma.ai sells the head unit (currently the comma 3X, with the comma four entering production) plus a vehicle-specific harness, and gives away the openpilot software and most schematics; revenue derives from device and accessory sales and not from a recurring software subscription.[^4][^9][^15] This model contrasts sharply with the multi-billion-dollar capital requirements typical of full-stack robotaxi developers such as [Waymo](/wiki/waymo) and Cruise, both of which build their own vehicles or modify them extensively, and with the model used by Tesla, which sells optional driver-assistance and "Full Self-Driving" software on cars whose hardware Tesla itself manufactures.[^11][^17] ## Products ### openpilot software openpilot is an open-source driving stack maintained primarily by comma.ai on GitHub. The project repository describes openpilot as "an operating system for robotics" and reports that openpilot "upgrades the driver assistance system on 300+ supported cars," with users contributing data totaling more than 100 million miles of driving as of 2025.[^3][^5] Documented supported brands span more than 40 manufacturers, including Acura, Audi, Chevrolet, Chrysler, Ford, Genesis, GMC, Honda, Hyundai, Jeep, Kia, Lexus, Lincoln, Mazda, Nissan, Ram, Rivian, Subaru, Toyota and Volkswagen Group brands (SEAT, Skoda, CUPRA).[^15] openpilot is released under the [MIT License](/wiki/mit_license); a handful of vendored components carry other permissive licenses where required by upstream authors.[^3] On supported vehicles the software replaces or augments the factory adaptive cruise control (ACC) and lane-keeping assist (often marketed as ALC or LKAS), and adds an automated lane-change function that initiates when the driver activates the turn signal and confirms the maneuver is safe.[^3][^17] #### openpilot architecture The runtime stack runs on a customized Ubuntu-based Linux distribution that comma.ai calls AGNOS, deployed onto the device's eMMC or NVMe storage; the head unit reads four imager-grade camera feeds (forward triple-camera array plus a driver-monitoring camera), an inertial measurement unit, GNSS receivers, the steering-angle sensor, and the host vehicle's CAN bus traffic.[^2][^3] These inputs are funneled into a single large [neural network](/wiki/neural_network) (referred to inside the project as "supercombo" through openpilot 0.9, and progressively replaced by newer end-to-end policies in later releases) whose outputs include the predicted future trajectory in three-dimensional space, lane-line and road-edge positions, the position and motion of a lead vehicle, and diagnostic probabilities such as the likelihood of driver disengagement.[^2][^22] The supercombo model historically combined a [convolutional](/wiki/convolutional_neural_network) image backbone derived from ResNet variants, a recurrent block (a GRU) for temporal context, and a set of fully connected output heads producing trajectories and auxiliary predictions; this end-to-end formulation places comma.ai broadly in the [imitation learning](/wiki/imitation_learning) lineage of autonomous-driving research, in which the policy is trained predominantly on logged human driving behavior rather than on hand-engineered cost functions.[^22] More recent openpilot releases have moved further toward pure end-to-end training, including end-to-end lateral planning in 2022 and end-to-end longitudinal control thereafter, with the network producing throttle and brake decisions directly from learned representations rather than from a separate model-predictive controller.[^23] Driver monitoring is performed by a second [neural network](/wiki/neural_network) that consumes the cabin-facing camera; the device pings the driver with escalating chimes and visual cues if attention is lost, and (after sustained inattention) disengages the system and brings the car to a controlled stop.[^24] On the comma two and later devices, near-infrared illuminators allow the driver-monitoring camera to work in the dark and through most polarized sunglasses.[^16][^24] #### openpilot safety model openpilot's safety architecture is intentionally decoupled from the neural network. The runtime model code runs on the device's Linux side, while a real-time microcontroller (the embedded "panda") sits between the device and the vehicle CAN bus and enforces a small set of hard limits.[^24][^25] comma.ai documents three principles: the driver must always be paying attention, the driver must always be able to instantly retake manual control, and the vehicle must not alter its trajectory faster than the driver can react.[^24][^25] In practice this means that: - Steering torque commands from openpilot are capped to a magnitude and slew rate small enough that an average driver can override them simply by holding the wheel; openpilot's documentation cites a maximum lateral deviation of about 1 m over 0.9 s.[^25] - The panda firmware blocks any acceleration request that would exceed conservative jerk and acceleration limits.[^25] - The brake pedal directly transitions the panda's "controls_allowed" state to false, immediately disengaging openpilot regardless of what the device-side neural network is requesting.[^24] - openpilot does not market itself as performing automatic emergency braking; it explicitly relies on the host vehicle's stock AEB and forward-collision warning where present.[^3][^24] The architecture means that even if the learned policy malfunctions, the safety envelope is preserved by independent firmware running on an entirely separate microcontroller, a design comma.ai has compared to safety-critical avionics in its blog posts.[^24] ### comma EON Dashcam DevKit (2017 to 2019, retired) The EON Dashcam DevKit, released October 19, 2017 at $699, was the first commercial openpilot device after the cancellation of the comma one. It was built around a OnePlus 3T smartphone running a custom Android-based operating system (NEOS) and shipped with a Panda CAN interface; the user was responsible for connecting the Panda to the host vehicle's CAN bus through a model-specific harness.[^2] The EON established the device-plus-harness retail pattern still used by comma.ai today.[^2] ### comma two (2020 to 2021, retired) Released January 7, 2020 at $999, the comma two replaced the EON. Externally it adopted a dashcam-style enclosure that attached to the windshield directly above the rearview mirror, was powered through USB-C from the harness rather than from a battery, and used a fan-based cooling loop sized for sustained sunlight on the windshield.[^16] The cabin-facing camera was monochrome with two infrared LEDs, enabling night-time driver monitoring without distracting the driver with visible illumination. A fourth CAN bus exposed through the OBD-II port allowed openpilot to read diagnostic streams that the factory ADAS module did not need to share with the device.[^16] ### comma three devkit (2021 to 2023) Announced July 31, 2021, the comma three was the first device that comma.ai designed and assembled in-house rather than basing on a third-party smartphone. The head unit integrates the Snapdragon 845 SoC underclocked to about 1.7 GHz, an Adreno 630 GPU used for [neural network](/wiki/neural_network) inference via the tinygrad-derived runtime, a 6-inch OLED display, and three road-facing cameras: a wide-angle camera, a main camera, and a tele camera with longer focal length.[^2][^18] Audio output and a microphone enable spoken alerts; the device runs an Ubuntu-based variant of AGNOS. The 256 GB devkit was sold for $1,999 and a 1 TB variant for $2,499.[^18] ### comma 3X (2023 to present) Released October 12, 2023, the comma 3X re-implements the comma three architecture with manufacturing and reliability as primary design goals. It retains the Snapdragon 845 compute, the 6-inch OLED display, and the three-camera vision system, but moves the panda function onto a single in-board STM32H7 microcontroller capable of CAN FD; upgrades the imagers to OX03C10 sensors with 140 dB HDR; and reduces the total part count by approximately half.[^9] The on-board storage is 128 GB.[^9] List price at launch was $1,250.[^9] ### comma four (announced 2025) The comma four was announced at COMMA_CON 2025 (with widely-shared documentation referring to a November 25, 2025 announcement) and is described by comma.ai as having the same compute and sensor suite as the comma 3X but in an enclosure roughly one-fifth the size, intended to hide entirely behind the rearview mirror with only a 1.9-inch, 300 PPI OLED touchscreen visible.[^21] List price is $999, or $699 with a trade-in of an older comma device.[^21] ### panda and red panda CAN interfaces The panda is a USB-to-CAN interface produced by comma.ai that exposes a vehicle's CAN buses to a host computer; it is sold both as a component of comma.ai's main devices and as a standalone product for car-hacking research.[^26] The standalone "red panda" is the high-end variant: an STM32H725-based device with CAN FD support across three buses plus a multiplexed bus, used both for diagnostics and for connecting a comma three to a CAN-FD-only vehicle that the in-board panda cannot natively reach.[^26] panda firmware and PCB designs are released openly on GitHub.[^26] ### tinygrad framework tinygrad is a minimalist deep-learning framework whose initial implementation was written by Hotz; it began as a teaching exercise but is now used in production as the inference engine for openpilot models running on the comma 3X's Adreno GPU.[^8][^22] After Hotz stepped down from operational duties at comma.ai in October 2022, he founded the tiny corp in November 2022 to develop tinygrad and a line of hardware appliances aimed at machine-learning training and inference workloads; in May 2023 the tiny corp announced a $5.1 million seed round.[^8] tinygrad is intentionally small (the project's repository advertises a strict line-count budget) and emphasizes portability across [Qualcomm](/wiki/qualcomm) mobile GPUs, AMD GPUs, and other accelerators.[^8] Both tinygrad and openpilot are listed as projects within the broader "comma" GitHub organization.[^3] ### comma body comma body is a wheeled robotics development kit, sold by comma.ai's online store, that runs openpilot in the same Ubuntu/AGNOS environment used on cars. Early body units were two-wheeled self-balancing platforms; later production units ship with helper wheels for static stability. From openpilot's perspective the body presents itself as a "car" over a CAN bus, including an OBD-C connector and hall-effect odometry sensors, so that the same logging and learning infrastructure used for road vehicles applies to the body's data as well. The chassis is powered by a 25.2 V, 100 Wh lithium-ion battery, and the device is intended as a research platform on which third parties can experiment with closed-loop robotics applications.[^27] ### Other community-facing activities comma.ai runs an annual conference, COMMA_CON, alongside community hackathons branded as comma hack day (sometimes numbered, e.g. Comma Hack 5) and online community programs such as the SDx Hacker Club.[^28] These events typically take place at the company's San Diego headquarters and have served as both recruiting funnels and release moments for new hardware, including the comma 3X (COMMA_CON 2023) and the comma four (COMMA_CON 2025).[^9][^21][^28] ## Funding Reported funding totals approximately $8.1 million across two disclosed rounds.[^10][^11] [Andreessen Horowitz](/wiki/andreessen_horowitz) led a $3.1 million seed in April 2016 at a $20 million pre-money valuation, with Chris Dixon joining the board.[^10] An April 2018 Series A filing reported an additional roughly $5 million at a valuation around $50 million; the lead investor in this round is not publicly named in the filings consulted here.[^11] In comparison with funding levels at full-stack autonomy companies, comma.ai has frequently been described as one of the most capital-efficient companies in the sector relative to fleet size and deployed software.[^11][^17] ## Research The most cited public research artifact from comma.ai is the 2016 arXiv paper "Learning a Driving Simulator" by Eder Santana and George Hotz (arXiv:1608.01230, submitted August 3, 2016), which trained a variational autoencoder and a generative-adversarial cost function on dashcam video from a Point Grey camera mounted in a 2016 Acura ILX, then learned a transition model in the embedded space using an action-conditioned recurrent [neural network](/wiki/neural_network).[^7] The paper released about 7.25 hours of driving data, divided across 11 videos, as the "comma.ai research" dataset.[^7] The paper has been cited in subsequent literature on driving world models and forms part of the broader chain of work that later companies such as [Wayve](/wiki/wayve) continued and scaled up.[^7] A separate body of academic literature has analyzed openpilot itself rather than being produced by comma.ai. The 2022 paper "Level 2 Autonomous Driving on a Single Device: Diving into the Devils of Openpilot" (arXiv:2206.08176) provides one such external analysis of the openpilot supercombo network and its training data assumptions, dissecting the convolutional, recurrent, and output-head architecture and discussing failure modes.[^22] The ACM paper "Finding Property Violations through Network Falsification" used openpilot's perception network as a target for formal-verification techniques.[^29] ## Adoption and reception ### Fleet and usage comma.ai publishes high-level fleet figures on its blog and in conference talks rather than continuously. The openpilot repository's README reports more than 300 supported cars and more than 100 million accumulated miles driven by openpilot users.[^3][^5] Community-shared statistics suggest that in a representative month around the comma 3X launch, roughly 1,200 cars logged about 1 million miles each month under openpilot; over time the supported vehicle count has grown from a handful at launch in 2017 to more than 300 in 2024 to 2025.[^5][^15] Real-world fleet uploads (with user consent) are the principal training-data source for openpilot's policy network, completing a closed loop between user devices, comma.ai's training infrastructure, and over-the-air model updates.[^2][^3] ### External evaluations In November 2020 Consumer Reports ranked the comma two and openpilot above every other driver-assistance system it had tested, including Tesla Autopilot, GM Super Cruise, and Ford Co-Pilot 360, citing driver-engagement, ease-of-use, and lane-centering quality. The comma two was the only aftermarket system in the test.[^17] The Insurance Institute for Highway Safety (IIHS) issued a 2024 evaluation of partial-automation systems that gave most stock automaker systems "poor" overall ratings; openpilot was not included in the IIHS test set, since the IIHS focuses on production-equipment configurations rather than aftermarket retrofits.[^30] A 2024 cross-country attempt covered by *Road & Track* and discussed by enthusiast outlets reported that a 2017 Toyota Prius running openpilot completed the New York to Los Angeles route in roughly 43 hours, with openpilot reportedly engaged for about 98% of the trip.[^2] ### Community and brand The openpilot ecosystem has an active third-party developer community that maintains "forks" of the software, ports it to additional vehicles, and tunes lateral and longitudinal behavior; the upstream project has historically been the dominant fork, but variants such as FrogPilot and sunnypilot have made small inroads with users who prefer alternative defaults.[^2] On forums such as Reddit and the discourse-hosted comma.ai community, the project is widely associated with enthusiast retrofit culture rather than mass-market consumer adoption.[^15][^28] ## Limitations and criticisms openpilot's stated capabilities are deliberately limited. The project's documentation classifies it as a Level 2 driver-assistance system, requires continuous driver attention, and disclaims any guarantee of safety or reliability; the GitHub README emphasizes that the software is "alpha quality" for research purposes only and ships with no warranty.[^3][^24] Like other camera-based [advanced driver-assistance systems](/wiki/adas), openpilot does not include radar or [LiDAR](/wiki/lidar) sensing of its own and depends on the host vehicle's stock automatic emergency braking (where present) to react to many imminent-collision scenarios.[^3][^24] Other limitations and criticisms commonly raised by reviewers and academics include: - **Aftermarket installation risk.** A 2016 Jalopnik analysis argued that the original comma one's installation process, in which the device tapped directly into a production vehicle's CAN bus, raised legitimate consumer-safety questions, and that NHTSA's special-order letter was justified given the lack of certification.[^14] - **No formal certification.** Unlike production-vehicle ADAS, openpilot is not certified to any specific safety standard, although comma.ai documents that the project's safety code is designed to follow ISO 26262 and MISRA C:2012 guidelines.[^24][^25] - **Camera-only sensor stack.** The vision-only architecture limits the system in heavy precipitation, dense fog, and low-light scenes where the host vehicle's headlights and the windshield's optical clarity dominate performance.[^17][^24] - **Fork fragmentation.** Community forks sometimes loosen the driver-monitoring or actuation limits in ways that contradict comma.ai's published safety design, which critics argue undermines the integrity of the safety model that comma.ai itself documents.[^24] - **Driver overconfidence.** Several reviewers have noted that high lane-centering quality, combined with the comma 3X's polished UI, can lead to misuse by drivers who interpret openpilot's smoothness as a sign of higher autonomy than the project actually claims.[^17][^24] ## Comparison The following table compares comma.ai with several adjacent driver-assistance and autonomy efforts as of late 2025. | System | Operator | Approach | Sensor stack | Geographic scope | Source availability | | --- | --- | --- | --- | --- | --- | | openpilot / comma 3X | comma.ai | Aftermarket retrofit, Level 2[^3] | 4 cameras + IMU + host CAN[^3] | 300+ supported cars worldwide[^3] | Open source, [MIT](/wiki/mit_license)[^3] | | Tesla Autopilot / FSD | Tesla, Inc. | OEM-installed, Level 2 (FSD beta)[^17] | 8 cameras (HW3/HW4) + IMU[^17] | Tesla cars in TSLA-supported markets[^17] | Closed source | | GM Super Cruise | General Motors | OEM-installed, Level 2[^17] | Camera + radar + LiDAR map[^17] | Specific GM models in NA[^17] | Closed source | | [Mobileye](/wiki/mobileye) SuperVision | Intel/Mobileye | OEM-supplied[^31] | 11 cameras + radar[^31] | OEM partners (e.g. Zeekr, Polestar)[^31] | Closed source | | [Waymo](/wiki/waymo) Driver | Waymo / Alphabet | Geofenced robotaxi, Level 4[^32] | Multi-modal: cameras, [LiDAR](/wiki/lidar), radar[^32] | Phoenix, San Francisco, Los Angeles, Austin (2025)[^32] | Closed source | | Cruise | General Motors | Geofenced robotaxi (paused 2023 to 2024)[^32] | Multi-modal[^32] | Limited U.S. cities[^32] | Closed source | comma.ai's design point is qualitatively different from both Tesla Autopilot (an OEM-installed Level 2 driver assistant on closed hardware) and [Waymo](/wiki/waymo) (a Level 4 robotaxi operator with custom multi-sensor stacks and high-definition mapping). It most closely resembles [Mobileye](/wiki/mobileye)'s historical positioning as a perception-and-control supplier, but with the deliberate inversion that comma.ai sells directly to end consumers as an aftermarket upgrade rather than to automakers as a Tier-1 component, and that comma.ai's software is open source.[^3][^31][^32] ## Related work The driving-research literature in which comma.ai's products and papers participate spans: - **End-to-end driving from imitation.** comma.ai's supercombo network and its open-source dataset are part of a tradition that includes [imitation learning](/wiki/imitation_learning) approaches developed at NVIDIA's PilotNet effort and at Waymo's ChauffeurNet group, and that more recently has been continued by [Wayve](/wiki/wayve) (LINGO and GAIA series) and by Tesla's vision-only stack.[^7][^22][^31][^32] - **Aftermarket retrofits.** The notion of upgrading a production vehicle's ADAS through an aftermarket device has been pursued, at various scales, by [Mobileye](/wiki/mobileye)'s consumer Mobileye 8 and 6 products as well as by smaller open-source projects, although openpilot remains the most widely deployed example.[^17][^31] - **Compute substrates.** openpilot inference on the Snapdragon 845's Adreno GPU is performed via tinygrad; this stands in contrast to most automotive ADAS deployments that rely on automotive-grade SoCs from [Mobileye](/wiki/mobileye) or on [NVIDIA Jetson](/wiki/nvidia_jetson_thor)-class boards.[^8][^31] - **Compact open-source toolchains.** tinygrad's minimalism stands alongside (and often interoperates with) [PyTorch](/wiki/pytorch) in the broader [machine learning](/wiki/machine_learning) ecosystem; comma.ai's choice to ship its own inference runtime is unusual among ADAS vendors of comparable size.[^8] ## See also - [Advanced Driver-Assistance Systems](/wiki/adas) - [Autonomous vehicle](/wiki/autonomous_vehicle) - [Mobileye](/wiki/mobileye) - [Waymo](/wiki/waymo) - [Wayve](/wiki/wayve) - [MIT License](/wiki/mit_license) - [Imitation Learning](/wiki/imitation_learning) - [Convolutional Neural Network](/wiki/convolutional_neural_network) - [Neural Network](/wiki/neural_network) - [LiDAR](/wiki/lidar) - [Qualcomm](/wiki/qualcomm) - [Andreessen Horowitz](/wiki/andreessen_horowitz) - [PyTorch](/wiki/pytorch) - [Machine Learning](/wiki/machine_learning) - [Deep Learning](/wiki/deep_learning) - [Computer vision](/wiki/computer_vision) - [Elon Musk](/wiki/elon_musk) - [NVIDIA Jetson Thor](/wiki/nvidia_jetson_thor) ## References [^1]: comma.ai, "comma.ai - make driving chill", comma.ai, 2025-10-15. https://comma.ai/. Accessed 2026-05-20. [^2]: Wikipedia contributors, "openpilot", Wikipedia, 2025-12-20. https://en.wikipedia.org/wiki/Openpilot. Accessed 2026-05-20. [^3]: commaai, "openpilot GitHub repository README", GitHub, 2025-11-15. https://github.com/commaai/openpilot. Accessed 2026-05-20. [^4]: comma.ai, "openpilot product page", comma.ai, 2025-10-15. https://comma.ai/openpilot. Accessed 2026-05-20. [^5]: comma.ai, "openpilot 0.9.8 release notes", comma.ai blog, 2024-04-30. https://blog.comma.ai/098release/. Accessed 2026-05-20. [^6]: comma.ai, "Introducing the comma 3X", comma.ai blog, 2023-10-12. https://blog.comma.ai/comma3X/. Accessed 2026-05-20. [^7]: Eder Santana and George Hotz, "Learning a Driving Simulator", arXiv:1608.01230, 2016-08-03. https://arxiv.org/abs/1608.01230. Accessed 2026-05-20. [^8]: George Hotz, "the tiny corp raised $5.1M", geohot.github.io, 2023-05-24. https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html. Accessed 2026-05-20. [^9]: comma.ai, "Introducing the comma 3X", comma.ai blog, 2023-10-12. https://blog.comma.ai/comma3X/. Accessed 2026-05-20. [^10]: Crunchbase News, "Automotive Automation Startup Comma Raises $5 Million In New Funding", Crunchbase News, 2018-05-09. https://news.crunchbase.com/venture/automotive-automation-startup-comma-raises-5-million-new-funding/. Accessed 2026-05-20. [^11]: comma ai (Medium), "A 100x Investment (Part 1)", Medium, 2019-03-04. https://comma-ai.medium.com/a-100x-investment-part-1-6c858bbe72ef. Accessed 2026-05-20. [^12]: Wikipedia contributors, "George Hotz", Wikipedia, 2025-12-18. https://en.wikipedia.org/wiki/George_Hotz. Accessed 2026-05-20. [^13]: Lucas Matney, "Comma.ai cancels the Comma One following NHTSA letter", TechCrunch, 2016-10-28. https://techcrunch.com/2016/10/28/comma-ai-cancels-the-comma-one-following-nhtsa-letter/. Accessed 2026-05-20. [^14]: Stef Schrader, "The Feds Were Right To Question The Safety Of The $999 Self-Driving Kit", Jalopnik, 2016-10-29. https://www.jalopnik.com/the-feds-were-right-to-question-the-safety-of-the-999-1788330916/. Accessed 2026-05-20. [^15]: commaai, "Supported Cars (docs)", openpilot docs, 2025-11-01. https://docs.comma.ai/CARS/. Accessed 2026-05-20. [^16]: Kyle Wiggers, "Comma.ai launches Comma Two, a $999 kit that imbues cars with assisted driving features", VentureBeat, 2020-01-07. https://venturebeat.com/ai/comma-ai-launches-comma-two-a-999-kit-that-imbues-cars-with-assisted-driving-features/. Accessed 2026-05-20. [^17]: Rob Stumpf, "Consumer Reports Ranks This Aftermarket Driver Assistance Kit Above Tesla Autopilot, Cadillac Super Cruise", The Drive, 2020-11-12. https://www.thedrive.com/news/37833/consumer-reports-ranks-this-aftermarket-driver-assistance-kit-above-tesla-autopilot-cadillac-super-cruise. Accessed 2026-05-20. [^18]: enjoi.dev, "Comma 3 Teardown, Inside The Comma Three Devkit", enjoi.dev, 2021-12-20. https://enjoi.dev/posts/2021-12-20-comma-3-teardown/. Accessed 2026-05-20. [^19]: Jay Traugott, "Comma AI Founder George Hotz Is Stepping Down, Company Still in Business", autoevolution, 2022-10-29. https://www.autoevolution.com/news/comma-ai-founder-george-hotz-is-stepping-down-they-are-still-in-business-202663.html. Accessed 2026-05-20. [^20]: Ingrid Lunden, "Hacker George Hotz, long a frenemy of Elon Musk, signs on for 12 weeks at Twitter", TechCrunch, 2022-11-21. https://techcrunch.com/2022/11/21/george-hotz-hacker-and-long-time-frenemy-of-elon-musk-signs-on-for-12-weeks-at-twitter/. Accessed 2026-05-20. [^21]: comma.ai, "Introducing the comma four", comma.ai blog, 2025-11-25. https://blog.comma.ai/comma-four/. Accessed 2026-05-20. [^22]: Li Chen et al., "Level 2 Autonomous Driving on a Single Device: Diving into the Devils of Openpilot", arXiv:2206.08176, 2022-06-16. https://arxiv.org/abs/2206.08176. Accessed 2026-05-20. [^23]: comma.ai, "End-to-end lateral planning", comma.ai blog, 2022-09-22. https://blog.comma.ai/end-to-end-lateral-planning/. Accessed 2026-05-20. [^24]: comma.ai, "Understanding the openpilot Safety Model", comma.ai blog, 2019-11-21. https://blog.comma.ai/understanding-the-openpilot-safety-model/. Accessed 2026-05-20. [^25]: comma.ai, "Safety (openpilot docs)", openpilot docs, 2025-11-01. https://docs.comma.ai/concepts/safety/. Accessed 2026-05-20. [^26]: commaai, "panda repository (red panda CAN interface)", GitHub, 2025-11-15. https://github.com/commaai/panda. Accessed 2026-05-20. [^27]: comma.ai, "Launching the comma body", comma.ai blog, 2022-11-14. https://blog.comma.ai/commabody/. Accessed 2026-05-20. [^28]: comma.ai, "COMMA_CON 2025 (ticketing)", tickets.comma-con.com, 2025-09-15. https://tickets.comma-con.com/events/comma/1859964. Accessed 2026-05-20. [^29]: Anneliese Day et al., "Finding Property Violations through Network Falsification: Challenges, Adaptations and Lessons Learned from OpenPilot", ACM/IEEE ASE 2022, 2022-10-10. https://dl.acm.org/doi/fullHtml/10.1145/3551349.3559500. Accessed 2026-05-20. [^30]: Insurance Institute for Highway Safety, "First partial driving automation ratings show industry doing poor job of preventing misuse", IIHS, 2024-03-12. https://www.iihs.org/news/detail/first-partial-driving-automation-ratings-show-industry-doing-poor-job-of-preventing-misuse. Accessed 2026-05-20. [^31]: Mobileye, "About Mobileye", Mobileye corporate site, 2025-10-15. https://www.mobileye.com/about/. Accessed 2026-05-20. [^32]: Waymo, "About Waymo", Waymo corporate site, 2025-10-15. https://waymo.com/about/. Accessed 2026-05-20.