Openpilot – Operating system for robotics

Original Article ↗ Hacker News Discussion ↗

Business model & openness

  • Commenters note comma’s relatively small team, minimal VC funding, profitability, and slower headcount growth as a refreshing contrast to hype-driven AI startups.
  • Core code is MIT-licensed with no copyright assignment; some models are shipped as ONNX files in the repo, though there’s disagreement over how “open” all models are.
  • Several argue that even if the company died, the open codebase reduces rug-pull risk, though proprietary training or models could still be a choke point.

Functionality & user experience

  • Multiple users say Openpilot significantly reduces fatigue on long drives and “makes driving chill,” especially on highways.
  • It’s described as notably better than many factory ADAS in lane keeping and phantom-braking behavior; some compare it favorably to Tesla Autopilot, but not to Tesla’s full FSD.
  • Forks like Sunnypilot/Frogpilot are mentioned; one popular mode keeps human control over pedals but automates steering.
  • Some users had issues with mismatched or premature “supported” vehicle listings but report straightforward refunds.

Legality, liability & insurance

  • The software is labeled alpha and “for research,” with liability explicitly on the driver.
  • Where it’s legal is unclear and jurisdiction-dependent; in parts of the EU, steering/throttle modifications may require homologation.
  • Concerns: potential extra punishment or insurance denial after a crash; others report insurers (including specific examples) accepting it as an aftermarket safety device.
  • Some worry about stored video being attractive to law enforcement and insurers.

Hardware, integration & compatibility

  • Works via CAN bus (and sometimes OBD-II), piggybacking existing lane-keeping and adaptive-cruise actuators.
  • One user reports a listed model/year not truly working at the time; support for that car was added later.
  • There is interest in retrofits for unsupported cars, with projects like “retropilot” mentioned but seemingly stale.

Safety, tech stack & standards

  • Concern about Python being involved in controlling heavy machinery; others counter that Python handles high-level logic/ML, while a separate “panda” microcontroller enforces safety limits in MISRA-C with ISO-related processes and extensive testing.
  • Another commenter criticizes the safety-marketing: citing specific ISO standards and 100% line coverage as weaker claims than they sound; what the actual safety model guarantees remains unclear.
  • It’s emphasized that this is Level 2 ADAS: always driver-supervised, no formal self-driving guarantees.

ADAS vs full self-driving & SAE levels

  • Several clarify that Openpilot, Tesla Autopilot/FSD, and mainstream systems (Honda, Toyota, Hyundai, etc.) are all SAE Level 2 when driver supervision is required.
  • Discussion around Waymo’s choice to avoid partial automation: some view gradual assistance as inherently risky due to complacency; others see that framing as partly marketing.
  • Debate about how meaningful SAE levels are to consumers; a narrow Level 3 system might be less practically useful than a robust Level 2.

“Operating system for robotics” vs ROS

  • Many react skeptically to calling Openpilot an “operating system for robotics,” saying it’s really an automotive driver-assistance stack, not a general robotics framework.
  • Some expected a ROS alternative and felt misled by the phrasing.
  • Supporters argue it has IPC/serialization and can run on a general-purpose robot platform from comma, positioning it as a ROS successor focused on end-to-end ML rather than traditional SLAM/motion-planning stacks.
  • Critics from a robotics perspective argue this is oversold marketing: replacing the broader robotics ecosystem and ROS’s standardization role is far from achieved, and traditional robotics stacks won’t disappear soon.

Community, tooling & ecosystem

  • The project uses Discord heavily; some appreciate immediacy, but others dislike that valuable technical knowledge gets buried and is not search-indexed.
  • There’s mention of a CTF-based onboarding for contributors.
  • Some commenters are philosophically opposed to paying to “babysit” an alpha system and supply data to a private company; others feel the comfort and safety gains justify the cost and effort.