Tesla: Failure of the FSD's degradation detection system [pdf]

Original Article ↗ Hacker News Discussion ↗

A new NHTSA investigation highlights failures in Tesla’s Full Self Driving (FSD) degradation detection, alleging the system often failed to recognize when cameras were impaired until just before crashes and sometimes lost track of vehicles ahead. Commenters debate whether Tesla’s camera‑only approach can ever match or exceed human drivers, especially in poor visibility or unusual “long tail” scenarios, and contrast it with LiDAR‑equipped systems like Waymo that are showing markedly better safety records. Many see FSD (marketed as “Supervised”) as overpromised and underperforming, raising broader questions about safety standards, liability, and the ethics of deploying partially autonomous systems on public roads.

Degradation Detection and Crash Concerns

  • Central issue: NHTSA notes FSD often failed to detect camera-visibility degradation or only did so moments before crashes.
  • Commenters see this as especially serious because the human supervisor cannot see the system’s internal confidence.
  • Some share anecdotes of FSD confidently driving toward unseen obstacles or losing track of lead vehicles in degraded conditions.

Behavior in Adverse Conditions

  • Several owners report FSD/AP shutting off entirely in heavy rain, fog, or snow, reverting to manual control.
  • Others note inconsistent behavior: sometimes alerts for dirt/sun on cameras, sometimes none for fog.
  • One describes an automatic “clean camera” wiper-fluid routine with no explicit warning that vision is degraded.
  • Some say speed is auto-limited in low visibility; others say they rarely see automatic disabling.

Camera-Only Approach vs. LIDAR and Other Sensors

  • Strong recurring critique: dropping LIDAR is framed as cost-cutting that sacrifices safety; many call it “shameful engineering.”
  • Supporters argue vision-only can work in principle, citing Tesla’s occupancy networks and improved HW4 performance.
  • Critics stress camera limitations vs. human eyes (dynamic range, low light, depth, glare) and lack of binocular, movable sensors.
  • “Wile E. Coyote attacks” (painted tunnel entrances, fake roads, puddle illusions) are raised as failure modes for camera-only.
  • Some ask why jurisdictions haven’t mandated or incentivized LIDAR-based systems.

FSD Capability, Safety, and “Supervised” Autonomy

  • Experiences vary: some say FSD handles ~97–99.9% of their driving and is often “better than me,” especially on newer hardware.
  • Others call “FSD (Supervised)” a scam: if constant human supervision is required, it isn’t truly self-driving.
  • Waymo is cited as handling 100% of driving in its ODD, highlighting the gap between “almost works” and fully driverless service.
  • Concerns that FSD’s crash statistics are skewed because it disengages in bad conditions.

AI Reasoning and Reliability

  • Broader debate on whether modern AI has robust logical/common-sense reasoning.
  • Some argue frontier models still fail basic physical reasoning and numerical/logical tasks, implying risk in edge driving cases.
  • Examples given: odd LLM failures, long-tail road events (e.g., animals or debris falling onto highways) that require novel reasoning.

Regulation, Reporting, and Recalls

  • NHTSA report described as “preliminary” and “vague”; some think discussion is premature.
  • Others counter that experts have warned about these issues for a decade; the report simply formalizes known risks.
  • Concern that Tesla’s internal data/labeling limitations may undercount FSD-related crashes.
  • Mention that Tesla is highly recall-prone relative to other automakers (per a linked article).

Product Positioning and User Experience

  • Disagreement on whether Tesla is still “premium”: many describe interiors and build quality as spartan or cheap vs. price.
  • Some argue a “premium” or expensive product with FSD should include the most comprehensive sensors (e.g., LIDAR).
  • Complaints about missing features (e.g., CarPlay) and inconsistent communication about camera cleanliness and limitations.

Meta: Polarization and Discussion Quality

  • Multiple comments note Tesla/Elon topics quickly become polarized, with both heavy upvoting and flagging of negative stories.
  • Some lament that much of the thread rehashes entrenched pro/anti-Tesla views rather than engaging with new specifics from the report.