2026-06-18 👁️ — views

AV Safety Data Comparison — NHTSA SGO Reports, Miles Per Crash, and Regulatory Readiness

NHTSA SGO crash data compared: Tesla FSD vs. Waymo incident rates, normalization caveats, and what the numbers mean for permit expansion.

Article 19 in the Physical AI Benchmark Series

This is the most quantitative article in the series. NHTSA’s Standing General Order (SGO) 2023 created the first mandatory, standardized crash database for automated vehicle systems. For the first time, safety data from Tesla FSD, Waymo, and other AV operators can be compared — imperfectly, but officially. This article works through what the data says, what normalization is required to read it honestly, and what it means for regulatory permit expansion.

Section 1 — NHTSA SGO 2023: What It Requires

NHTSA’s Standing General Order, effective June 2023, requires all manufacturers of vehicles with SAE Level 2 or higher automation to report:

• Any crash resulting in a fatality, hospitalization, airbag deployment, or injury requiring medical attention
• Any crash involving a vulnerable road user (pedestrian or cyclist)
• Reports must be filed within 30 days of the crash (or 24 hours if the crash involved a fatality)

The SGO covers all AV operators — consumer L2 systems like Tesla FSD and Autopilot, and fully driverless L4 commercial fleets like Waymo Driver. It is the first mandatory, comparable crash database for automated driving systems in the US.

Critical normalization caveat — this is the most important framing in the article:

Tesla reports ALL miles driven under L2 automation (FSD and Autopilot), including hundreds of millions of supervised consumer miles where a licensed human driver is present and responsible. Waymo reports only commercial driverless operations, with no human supervisor in the vehicle. These are fundamentally different operational conditions. Per-mile comparisons between Tesla and Waymo require this normalization to be meaningful. A raw crash count comparison — without accounting for miles driven, automation level, and the presence or absence of a human supervisor — is not a valid safety comparison.

Section 2 — Published NHTSA SGO Data Summary

The following table summarizes publicly reported NHTSA SGO data through 2024. All figures are drawn from public NHTSA SGO disclosures and publicly available company safety reports. Estimates are labeled as such.

Per-mile rate comparison (publicly reported data, normalized):

• Tesla FSD: approximately 2+ crashes per million miles (L2, human supervisor present, based on NHTSA SGO data)
• Waymo: approximately under 1 crash per million miles (L4, fully driverless, based on NHTSA SGO data)
• US human driver baseline: approximately 4.2 crashes per million miles (NHTSA baseline)

Both Tesla FSD and Waymo report rates below the US human driver baseline. However, the comparison is not apples-to-apples: Tesla’s L2 rate includes supervised consumer driving (human in loop, responsible for the vehicle); Waymo’s L4 rate is fully driverless commercial operation. A lower crash rate for a supervised L2 system is expected — the human driver is the primary safety layer. The more meaningful comparison for AV safety is the L4 driverless rate against the human baseline.

Section 3 — What the Data Means for Regulatory Expansion

Regulators — CA DMV, NHTSA, and state-level AV agencies in Arizona, Texas, and Nevada — use safety data as the primary gate for permit expansion decisions.

Waymo’s case for expansion:

Waymo’s 0 confirmed driverless fatalities across 30M+ commercial miles is the strongest safety record of any commercial AV operator globally as of mid-2026. The CA DMV and AZ DMV have both cited this record in permit decisions. The data supports Waymo’s current multi-state expansion strategy — San Francisco, Phoenix, Los Angeles, and Austin — because the safety record provides regulators with a statistically meaningful dataset to evaluate.

Tesla’s regulatory path:

Tesla’s SGO data is complicated by the L2 vs. L4 distinction. Regulators require L4 (no human supervisor) safety data before issuing driverless commercial permits. Tesla’s 500M+ miles of L2 FSD data demonstrates software capability under human supervision — but does not directly address the regulatory question of whether FSD is safe with no human in the loop. Tesla has not yet accumulated significant L4 driverless commercial miles in a form that satisfies the regulatory evaluation framework for driverless permits. This is the single largest data gap in Tesla’s regulatory path as of mid-2026.

The Cruise lesson:

The October 2023 pedestrian drag incident — in which a Cruise vehicle failed to stop after striking a pedestrian and dragged the injured person approximately 20 feet — was not simply a safety failure. The incident response was the regulatory trigger. Cruise did not promptly and fully disclose the incident to CA DMV. The permit suspension that followed in November 2023 was driven as much by the inadequacy of the incident response as by the incident itself. Waymo responded to Cruise’s suspension by publishing more detailed safety transparency reports. The lesson for all AV operators: regulator trust is built incrementally through data transparency, and lost rapidly through any perception of concealment.

Section 4 — Incidents Per Million Miles Trend (Waymo Estimates)

The following table shows estimated improvements in Waymo’s commercial driverless incident rates over time. Waymo does not publish precise per-mile incident rates; these estimates are derived from publicly available safety reports, SGO data, and industry analyst interpretations. All figures are estimates.

The trajectory implied by publicly available data is consistent improvement over time, correlated with software maturity and accumulated operational experience in each city. This trend is the core quantitative argument Waymo makes to regulators when applying for permit expansions.

Section 5 — Safety as the Regulatory Currency

The NHTSA SGO framework has created a de facto safety data marketplace: companies that accumulate more driverless miles with lower incident rates earn regulatory trust faster. The following table summarizes the current state.

The quantitative standard emerging from SGO data:

No regulator has published an explicit safety threshold — a “crashes per million miles” number below which permits are automatically granted. But the trajectory of CA DMV decisions implies an informal standard: operators with 0 driverless fatalities and a downward incident trend over a substantial mileage base (10M+ commercial driverless miles) are approved for expansion; operators with unresolved incidents or inadequate transparency are not.

The SGO data is imperfect — voluntary in some dimensions, inconsistently formatted across operators, and subject to normalization problems (the L2 vs. L4 issue). But it is the only mandatory, standardized safety database the US has for AV systems, and it is now the primary evidentiary basis for permit decisions. Companies that invest in safety data quality and transparency — not just safety performance — are building the regulatory asset that determines their expansion speed.

How This Article Fits the Series

This is article 19 in the Physical AI Benchmark Series. The series has covered:

• Articles 1–9: Technology, regulation, capital, and the master scorecard
• Articles 10–13: Four supply-side structural constraints (HD mapping, teleop, OTA, FMVSS)
• Article 14: Updated scorecard integrating all four constraints
• Article 15: The demand side — rider experience, adoption curves, and pricing
• Article 16: The supply chain — manufacturing partners, fleet operations, and distribution ecosystem
• Article 17: Investment-grade competitive moat analysis — durable vs. temporary advantages
• Article 18: Tesla Cybercab vs. Model Y robotaxi — two vehicles, two timelines, one ramp
• Article 19 (this article): AV safety data — NHTSA SGO reports, miles per crash, and regulatory readiness

Sources

• NHTSA Standing General Order crash data — NHTSA ↗
• Waymo safety report — Waymo ↗
• Tesla FSD safety disclosures — Tesla ↗
• California DMV AV permit decisions — CA DMV ↗

Tags