How Close Are We to Self-Driving Cars, Really?
/ 8 min read
Table of Contents
The gap between today’s best self-driving systems and a steeringless car you’d trust with your child remains enormous — but one company is closing it faster than most realize. Waymo has driven 127 million fully driverless miles through September 2025 with 90% fewer serious injury crashes than human drivers, validated by peer-reviewed studies and independent insurance data from Swiss Re. Tesla FSD, despite viral cross-country drives, still operates under human supervision with an intervention rate roughly 460x worse than what its own engineers say is needed for unsupervised operation. No company has received federal approval to mass-produce a steeringless consumer vehicle, and credible forecasts place that milestone no earlier than 2035–2040.
The autonomous driving revolution is real — but the “nines of reliability” math reveals just how far the hardest part of the journey still extends.
The metrics that actually matter
The autonomous vehicle industry suffers from a measurement crisis. Companies cherry-pick favorable statistics, definitions vary wildly, and the most commonly cited metric — California DMV disengagement reports — is widely misunderstood.
Miles per disengagement, reported annually to the California DMV, captures how far an AV drives before a human safety driver takes over. In the 2024 reporting period, 31 companies operated 2,819 vehicles across 6.3 million test miles. Waymo logged roughly 29,000+ miles per disengagement in its best reporting years, while new entrants like May Mobility managed just 0.66 miles. The DMV itself warns these reports are “not intended to compare one company with another” — companies define disengagements differently, safety drivers vary in caution, and the reports exclude commercial operations. California plans to replace this metric entirely in 2026.
A deeper problem is distinguishing critical interventions (where failure to act would cause a crash) from precautionary takeovers (where a cautious driver grabs the wheel unnecessarily). A 2025 peer-reviewed analysis found 77% of all disengagements were initiated by safety drivers, not by AV system failures.
The RAND Corporation’s landmark study exposed an even more fundamental challenge: to prove with 95% confidence that an AV has a fatality rate just 20% lower than humans, it would need to drive approximately 8.8 billion miles — roughly 400 years with 100 vehicles operating 24/7. This is why RAND researchers concluded “developers cannot simply drive their way to safety.”
What the “nines” framework reveals
Borrowed from telecommunications reliability engineering, the “nines” framework translates autonomous driving performance into an intuitive scale. Applied to crash-free miles, the math is unforgiving.
| Reliability | Crash-free rate | One crash every… | Real-world meaning |
|---|---|---|---|
| 99% (2 nines) | 99% | 100 miles | Multiple crashes per week |
| 99.9% (3 nines) | 99.9% | 1,000 miles | ~14 crashes/year for average driver |
| 99.99% (4 nines) | 99.99% | 10,000 miles | ~1.4 crashes/year — still unacceptable |
| 99.999% (5 nines) | 99.999% | 100,000 miles | ~1 crash every 7 years |
| 99.9998% (5.7 nines) | 99.9998% | ~530,000 miles | Average human driver |
| 99.9999% (6 nines) | 99.9999% | 1,000,000 miles | ~2x safer than humans |
| 99.99999% (7 nines) | 99.99999% | 10,000,000 miles | ~19x safer than humans |
| 99.9999988% (~8 nines) | 99.9999988% | ~86,000,000 miles | Human fatal crash rate |
Using 2023 NHTSA data: Americans drove 3.25 trillion miles and experienced 6.14 million police-reported crashes, yielding one crash per ~529,000 miles — roughly 5.7 nines. For fatal crashes, the rate is one per ~86 million miles — approximately 8 nines.
A critical caveat: NHTSA estimates 60% of property-damage crashes and 32% of injury crashes go unreported, which would push the true all-crash rate to roughly one per 262,000 miles. AVs, by contrast, report virtually every incident, creating an inherent reporting asymmetry that makes direct comparisons deceptively unfavorable to autonomous systems.
Waymo: the clearest proof it works
Waymo’s data represents the strongest evidence that autonomous vehicles can meaningfully outperform human drivers — validated by peer-reviewed studies, Swiss Re insurance claims analysis, and mandatory NHTSA crash reporting.
Through September 2025, Waymo accumulated 127 million rider-only miles across Phoenix, San Francisco, Los Angeles, and Austin. The company now completes 450,000+ paid rides per week across 10 U.S. cities, accumulating roughly 2 million autonomous miles weekly. Every one of these miles is driven with no human behind the wheel.
The safety results, published in Traffic Injury Prevention across 56.7 million rider-only miles, show statistically significant improvements:
- 90% fewer serious-injury-or-worse crashes (0.02 vs. 0.23 per million miles)
- 81% fewer injury crashes (0.74 vs. 3.97 per million miles)
- 82% fewer airbag-deployment crashes
- 92% fewer pedestrian injuries
- 83% fewer cyclist injuries
The Swiss Re insurance study (December 2024) analyzed 25.3 million autonomous miles and found an 88% reduction in property damage claims and 92% reduction in bodily injury claims — even when compared exclusively to newer vehicles (2018–2021) with advanced driver-assistance systems.
Waymo is not without failures. NHTSA opened an investigation after 19–20 documented instances of Waymo vehicles passing stopped school buses in Austin. A San Francisco power outage caused some vehicles to freeze in intersections. But the overall record across 127 million miles remains substantially better than human driving by every measured metric.
Tesla FSD: rapid progress meets hard limits
Tesla’s Full Self-Driving tells two very different stories depending on which data you examine.
The optimistic narrative centers on FSD v14 (released October 2025). Crowdsourced data shows v14 achieving roughly 1,454 miles between critical disengagements — a 2.7x improvement over v13 and 8x improvement over v12.5. On December 31, 2025, a Tesla enthusiast completed the first documented zero-intervention coast-to-coast drive — 2,732 miles from LA to Myrtle Beach on FSD v14.2.
The sobering reality emerges from independent testing. AMCI Testing recorded one intervention every 13 miles on FSD v12.5. Tesla’s Austin robotaxi fleet has experienced 14 crashes in approximately 800,000 paid miles — 4x worse than an average Tesla driver without Autopilot and roughly 9x worse than the national average.
Tesla’s own head of FSD has stated the target for unsupervised operation is approximately 670,000–700,000 miles between critical interventions. At v14’s crowdsourced rate of 1,454 miles per critical disengagement, Tesla is still roughly 460x short of that target.
Elon Musk predicted coast-to-coast autonomous driving by end of 2017 (achieved December 2025, 8 years late), “over 1 million robotaxis” by 2020 (Tesla operates ~31 in Austin as of February 2026), and “millions of Tesla robotaxis” in the second half of 2025.
Where each company stands
| System | Best metric | Miles per event | Approx. “nines” | vs. human avg |
|---|---|---|---|---|
| Tesla FSD v12.5 (AMCI) | Miles/intervention | ~13 | 1.1 nines | ~40,000x worse |
| Tesla FSD v14 (crowdsourced) | Miles/critical disengagement | ~1,454 | 3.2 nines | ~364x worse |
| Tesla Austin robotaxi | Miles/crash | ~57,000 | 4.8 nines | ~9x worse |
| Waymo (CA DMV, 2020) | Miles/disengagement | ~29,000 | 4.5 nines | ~18x worse* |
| Waymo (police-reported crashes) | Miles/crash | ~476,000 | 5.7 nines | ~ Human avg |
| Waymo (injury crashes) | Miles/injury crash | ~1,350,000 | 6.1 nines | ~3x better |
| Waymo (serious injury) | Miles/serious crash | ~50,000,000 | 7.7 nines | ~10x better |
| Human average (all) | Miles/crash | ~529,000 | 5.7 nines | Baseline |
| Human average (fatal) | Miles/fatal crash | ~86,000,000 | 7.9 nines | Baseline |
*Disengagement and crash metrics are not directly comparable.
What each level means for a driver covering 14,000 miles per year:
- 3 nines: One event every ~1,000 miles — 14 events per year
- 5 nines: One event every ~100,000 miles — one event every 7 years
- 5.7 nines: One crash every ~530,000 miles — one crash every 38 years (human average)
- 7 nines: One crash every 10,000,000 miles — one crash every 714 years
The road to removing the steering wheel
No company has received federal approval to mass-produce a steeringless vehicle for public roads.
NHTSA’s March 2022 rule update replaced “steering wheel” with technology-neutral language, but covered only a fraction of relevant safety standards. The 2,500-vehicle exemption cap under federal law remains the binding constraint — no manufacturer can deploy more than 2,500 vehicles per year that don’t meet all existing standards. The SELF DRIVE Act, which would raise this cap, has failed to pass Congress for nearly a decade.
Zoox (Amazon) achieved the most significant milestone: in August 2025, it received the first-ever NHTSA exemption for an American-built steeringless autonomous vehicle — but only for 64 vehicles for demonstration purposes. GM Cruise abandoned its steeringless Origin entirely in 2024 after a pedestrian-dragging incident and a $583 million write-down. Waymo sidesteps the question by using modified production vehicles that retain steering wheels.
A McKinsey survey of 91 industry experts (January 2026) found consensus: large-scale L4 robotaxi rollout won’t arrive until ~2030. L4 capability in privately owned vehicles is expected around 2032, with fewer than 6% of vehicles sold in 2035 having any L4 functionality. Level 5 — true, go-anywhere autonomy — has zero representation in any credible forecast before 2035.
S&P Global Mobility states: “The ability for a consumer to buy a car that will drive itself everywhere without the driver ready to take the wheel is unlikely to happen by 2035.”
The math doesn’t lie
Three insights cut through the hype.
The “nines” framework exposes a logarithmic challenge. Each additional “nine” requires roughly a 10x improvement. Tesla FSD v14’s best performance (~3.2 nines) needs to improve by roughly 360x to match average human driving (~5.7 nines). The viral cross-country drives represent best-case highway performance — the easiest driving domain.
Waymo has quietly crossed the human-performance threshold on crash metrics — the first autonomous system to do so at meaningful scale with independently validated data. But it operates within carefully mapped geofenced areas, doesn’t handle snow, and employs remote operators. It has solved a constrained version of the problem exceptionally well.
The gap between robotaxi-in-a-geofenced-city and steeringless-car-for-everyone is measured in decades, not years. Federal regulations still require steering wheels. Comprehensive legislation has failed for a decade. Snow, construction zones, and unmapped areas remain unsolved at scale. Every credible forecast places mass-market steeringless consumer vehicles no earlier than 2035–2040.
The autonomous future is coming. It’s just coming on a timeline measured in “nines” — and each one takes longer than the last.