What Is a Disengagement, and Why Does It Matter?
A disengagement, in the context of California's autonomous vehicle testing regulations, is a specific regulatory event: the deactivation of the autonomous mode while the vehicle is operating on public roads, either initiated by the safety driver (who judged the system was about to operate incorrectly or unsafely) or by the automated driving system itself (which detected a situation it could not handle and requested human intervention). Companies testing autonomous vehicles on California public roads are required to file annual reports documenting these events.
The metric derived from these reports — miles per disengagement, or MPD — has become the most widely cited quantitative indicator of autonomous driving maturity. The logic is intuitive: if a system requires human intervention every few miles, it is clearly not ready for commercial deployment; if it can operate thousands of miles between interventions, it is approaching operational readiness. The reality, as with most quantitative metrics in complex domains, is considerably more nuanced.
The California DMV Mandate: What Companies Must Report
California's Autonomous Vehicle Testing Regulations, administered by the Department of Motor Vehicles, have required autonomous vehicle testing companies to file annual disengagement reports since 2015. The reports must cover all testing on California public roads and include: the total miles driven in autonomous mode, the number of disengagements, the location and date of each disengagement, and a brief description of the circumstances that led to it.
The definition of a reportable disengagement has been refined over the years but remains somewhat subjective: a disengagement caused by a "failure of the autonomous technology" or a "situation the system was not designed to handle" must be reported. Disengagements for comfort reasons — a safety driver who prefers not to use the system in a particular road segment, or who takes manual control as a precautionary measure rather than due to a specific system limitation — are technically not required to be reported, though many companies report all disengagements regardless of cause in the interest of comprehensive data.
The Headline Numbers: A Story of Dramatic Progress
The aggregate trajectory of California disengagement data tells a compelling story of technological improvement. Waymo, consistently the leader in miles per disengagement, reported approximately 13,000 miles per disengagement in its 2019 filing. By its 2022 filing, that figure had grown to over 1 million miles per disengagement across its driverless fleet — a number so large it stretches the concept of a meaningful disengagement rate because the fleet is operating commercially without safety drivers in a growing portion of its miles.
Other companies showed similar trajectories, though at lower absolute values. Cruise reported approximately 28,000 miles per disengagement in 2022 before its operational suspension in late 2023. Zoox, Aurora, and Nuro all demonstrated consistent year-over-year improvement in their respective test programs. Even companies with relatively modest mile counts showed order-of-magnitude improvements in intervention rates compared to their earliest filings.
"A system that disengages once per million miles is not a system that is failing — it is a system that is monitoring itself with extraordinary precision and flagging the exact scenarios it has not yet mastered."
Why Cross-Company Comparison Is Deeply Misleading
The fundamental problem with using disengagement data to compare companies is that different companies test in different operational domains, under different definitions of "disengagement," and with different organizational cultures around what constitutes a reportable event. A company testing exclusively in a geofenced suburban environment with wide lanes and light pedestrian traffic will naturally achieve higher miles per disengagement than one testing in dense urban cores with complex intersections and unpredictable pedestrian behavior. The numbers reflect the difficulty of the environment as much as the capability of the technology.
The definition of a safety-critical disengagement versus a precautionary one varies substantially between companies. Waymo has historically reported only disengagements that would have resulted in a safety-relevant outcome if the safety driver had not intervened — a more restrictive definition that tends to produce higher MPD numbers. Other companies report all manual takeovers, including those taken for comfort or convenience. Without standardized definitions, direct comparison of absolute MPD numbers between companies is an exercise in comparing apples to oranges.
Waymo's Data Trajectory: A Case Study
Waymo's disengagement reports are the most instructive in the industry not because of their headline MPD numbers — which are inflated by the company's shift to driverless operation — but because of the granular incident descriptions that accompany them. Each disengagement report includes a description of the scenario that triggered the intervention, creating a qualitative audit of the system's remaining limitations.
Reading through successive years of Waymo's reports reveals a consistent pattern: early reports are dominated by basic perception errors (failure to detect static obstacles, incorrect classification of vehicle types), which are gradually replaced by more subtle failures (misinterpretation of ambiguous pedestrian intent, difficulty with unusual vehicle movements). This evolution in failure mode character — from basic to subtle — is itself a meaningful indicator of maturity: a system that fails in complex, ambiguous situations is further along than one that fails in routine, clearly structured scenarios.
Beyond California: Other Safety Data Sources
California DMV disengagement reports are not the only source of publicly available AV safety data, though they are the most structured. NHTSA's Standing General Order, implemented in 2021, requires automatic reporting of any crash involving an SAE Level 2–5 automated driving system that results in property damage or personal injury within 30 seconds of the system being active. This data provides a different lens — focused on outcomes rather than system interventions — and covers all US states, not just California.
Individual companies publish safety reports of varying depth. Waymo's Waymo Safety Report, published annually, includes detailed statistical comparisons between Waymo's crash rate and national human driver benchmarks across comparable road types and conditions. Cruise, Zoox, and Aurora publish similar documents. These voluntary reports, while inevitably curated by their authors, provide richer context than the sparse regulatory filings and include internal metrics such as scenario coverage rates, simulation test results, and safety case arguments that are not required by any current regulation.
What Actually Matters: Safety at Fleet Scale
The ultimate metric for autonomous vehicle safety is not miles per disengagement — it is injury and fatality rate per billion vehicle miles traveled, compared against human driver baselines on equivalent road types. Disengagement rates are a proxy for this metric during development, when the fleet is too small and geographically constrained to produce statistically meaningful safety outcome data. But as fleets scale, the proxy becomes less necessary because the direct measurement becomes possible.
Waymo's published safety data, comparing its crash rate in San Francisco against NHTSA baseline data for similar urban driving conditions, shows a statistically significant reduction in injury-causing crash rate per mile relative to human drivers. These are early results from a relatively small fleet, and they do not yet constitute proof that AV systems are definitively safer than the average human driver under all conditions. But they represent a trajectory — and a methodology — that will become increasingly significant as fleets grow and data accumulates over the coming years.
The disengagement report era of AV safety measurement is gradually being superseded by something more meaningful: direct safety outcome comparisons at fleet scale. The industry's best practitioners have always known that miles per disengagement was a crude early metric. What replaces it — rigorous, standardized, independently audited safety outcome data — will be the definitive evidence that determines whether autonomous vehicles become the safety revolution their proponents claim they can be.