Frequently Asked Questions
Everything you need to know about our methodology, data sources, incident definitions, and how to interpret robotaxi safety statistics
How we calculate miles per incident and where our data comes from
NHTSA SGO 2021-01 is a federal requirement that mandates manufacturers and operators of vehicles equipped with SAE Level 2 ADAS or Levels 3-5 ADS to report certain crashes to NHTSA.
Key reporting requirements:
This is the primary public data source for tracking autonomous vehicle incidents across all manufacturers, including Tesla, Waymo, Cruise, and others.
MPI = Fleet Miles Driven ÷ Number of Incidents
We calculate this for each interval between consecutive incidents:
This gives us the actual distance driven between each incident, accounting for fleet growth over time.
Tesla disclosed in their Q3 2025 earnings report that their Austin robotaxi fleet was averaging approximately 115 miles per vehicle per day. This is consistent with typical robotaxi utilization rates (compared to ~30 miles/day for private vehicles).
Limitations:
Even with some uncertainty in this estimate, the exponential improvement trend (R² = --) is robust to reasonable variations in daily mileage assumptions.
Austin is the only location where Tesla operates unsupervised Level 4 autonomous driving. This is the only deployment that:
The Bay Area fleet operates with safety drivers (Level 2) as required by California law. Different reporting requirements apply, and including that data would conflate different operational modes.
We track daily fleet size using multiple sources:
The current Austin fleet is approximately -- unsupervised vehicles.
Understanding when incidents are reported and how this affects the data
Reporting lag is the delay between when an incident occurs and when it appears in NHTSA's public database. Under SGO 2021-01:
Why it matters: If you count "miles driven through Sunday" but incidents haven't been reported yet, you may be calculating MPI with an incomplete incident count, making safety appear better than it actually is.
This tracker displays both "Data Through" (latest mileage date) and "Latest Report" (most recent incident report date) so you can assess data freshness.
Current data status:
We check NHTSA's SGO database daily and update incident data as new reports are published. The "current streak" metric (miles without an incident) is inherently provisional until sufficient time has passed for any incidents to be reported.
The current streak (-- miles) should be treated as provisional, especially for the most recent 7-14 days.
Guidance:
The completed interval MPI values (between known incidents) are the most reliable metrics, as both the start and end incidents have been reported.
Different sources may show different incident counts due to:
This tracker specifically counts only Austin (unsupervised) incidents from NHTSA SGO ADS reports. Our methodology is documented and consistent.
Why "robotaxi vs human" comparisons are more complicated than they appear
Comparing robotaxi incident rates to human crash rates involves several mismatches:
These factors can make comparisons unfair in either direction, depending on what you're measuring.
Several benchmarks exist, each measuring something different:
| Baseline | MPI | What it measures |
|---|---|---|
| Police-reported crashes | ~500,000 | Crashes serious enough for police report |
| Insurance claims | ~300,000 | Crashes where insurance was involved |
| All crashes (estimated) | ~100,000-200,000 | Including unreported minor incidents |
| Urban taxi/rideshare | Unknown | Most relevant but not publicly tracked |
The insurance claim baseline (~300,000 MPI) is arguably the most appropriate comparison for robotaxis since SGO captures incidents that would result in claims.
An ideal comparison would control for:
Unfortunately, no public dataset exists for "Austin urban taxi/rideshare incident rate using SGO-equivalent reporting." The closest would be comparing to Uber/Lyft incident rates in Austin, but that data isn't public.
What we can say: Tesla's trend shows rapid improvement. Whether the current rate is "better" or "worse" than humans depends heavily on what comparison you choose.
Waymo reports approximately 1,000,000+ miles per incident based on their published safety data and the Swiss Re study.
Tesla's current MPI (-- miles) is approximately --x less than Waymo's reported figures.
Important caveats:
This is a fair criticism. Tesla's Austin robotaxis operate in a limited Operational Design Domain (ODD):
National human crash statistics include:
The geofence cuts both ways: Urban areas have more pedestrians, cyclists, and complex intersections than highways. The "easy" vs "hard" comparison isn't straightforward.
What counts as an incident and who's at fault
An "incident" on this site means a crash event that appears in NHTSA SGO 2021-01 ADS reports for Tesla's Austin robotaxi program.
This includes:
This does NOT include:
Yes. SGO 2021-01 has a lower reporting threshold than police reports. It can capture:
This is a legitimate criticism of direct MPI comparisons. A robotaxi might have a "worse" MPI than police-reported human stats while actually being safer, simply because more incidents are being reported.
Counterpoint: SGO doesn't capture all minor incidents — only those meeting specific criteria. Many true minor contacts may still go unreported.
SGO reports do not include fault determination. This is a significant limitation of the public data.
What we know:
Why this matters: If a significant portion of incidents are other vehicles hitting the robotaxi, the "true" at-fault MPI would be much better than the overall MPI.
We are exploring ways to provide provisional fault categorization based on available evidence, but this is inherently uncertain.
Based on available SGO data for Tesla's Austin fleet:
The lack of serious injuries or fatalities is notable, though the sample size is still relatively small. This suggests that even when incidents occur, they tend to be low-severity.
Yes. Tesla redacts the narrative section of SGO crash reports, preventing the public from knowing specific details about how crashes occurred.
What is NOT redacted:
This is a legitimate transparency concern. Waymo, by comparison, provides more detailed incident descriptions. However, the core metrics tracked here (MPI, trend) are calculable from non-redacted data.
Understanding the role of safety monitors and remote operators
The situation is mixed:
Tesla has been progressively removing safety monitors from Austin vehicles as confidence increases. The exact split is not publicly disclosed.
This is a key confound. If safety monitors intervene to prevent crashes, the observed MPI reflects "FSD + human backup" rather than "FSD alone."
What we don't know:
As Tesla removes more monitors, the data will increasingly reflect true autonomous performance. If MPI stays strong with fewer monitors, that's a positive signal.
Tesla has remote operators who can assist vehicles that get stuck or need guidance. However:
The extent to which remote assistance affects safety performance is unknown.
Fleet-wide removal of safety monitors is expected when:
This tracker focuses on the quantitative signals that predict when removal becomes operationally rational.
How to read the charts and understand the statistical methods
We use a logarithmic (log) scale because:
What to watch for: On a log scale, linear improvement looks like it's slowing down. Our trend line shows the data is consistent with exponential (not linear) improvement.
The trend line is an exponential regression of the form:
MPI = a × e^(b × days)
This model assumes safety improves at a constant percentage rate over time (like compound interest). Key outputs:
An R² close to 1.0 means the exponential model explains most of the variation in the data.
Yes, somewhat. With a small number of data points (currently -- incidents), each new incident affects the trend.
How we address this:
As more incidents occur, the trend becomes more robust. Early-stage data naturally has higher uncertainty.
Some people argue log scales "compress" the gap between Tesla and human baselines. This is both true and misleading:
We show the numeric values clearly. Tesla's latest MPI is -- miles. Human baseline (police-reported) is ~500,000 miles. You can judge the gap yourself.
The exponential improvement trend could break if:
We update daily and will note any significant deviations from the trend. If MPI stops improving exponentially, the doubling time estimate will increase or become undefined.
Addressing frequent misunderstandings and media claims
These reports are mathematically correct but contextually incomplete:
What they miss:
The question isn't "what was Tesla's average safety?" but "what is Tesla's current safety and trajectory?"
Different numbers arise from:
This tracker uses consistent methodology: Austin unsupervised ADS crashes from NHTSA SGO, updated daily.
This site aims to be data-driven and transparent:
What the data shows: Tesla's robotaxi safety is improving exponentially but currently lags behind Waymo and human baselines. We present both the positive trend and the current gap.
If you find errors or methodological issues, please report them on GitHub.
This is a personal decision that depends on your risk tolerance and how you weigh different factors:
We provide the data; you decide what it means for you.
Track Tesla robotaxi safety metrics in real-time with our interactive dashboard
View Safety Tracker