At the current stage, the takeover rate appears to have become the primary metric for evaluating autonomous driving capabilities. Automakers compete at launch events over "zero takeovers throughout the journey," third-party reviews rank vehicles by the number of takeovers per 100 km, and users habitually ask whether a car's autonomous driving is reliable when choosing a vehicle.
However, blindly pursuing a low takeover rate tends to overlook a critical point.
The takeover rate measures more of a result that can be strategically manipulated rather than the system's ability to make correct decisions in real-world scenarios. When R&D direction is driven by this single number, the optimization target easily shifts from "driving well" to "minimizing the metric."
01. What Exactly Is the Takeover Rate?
A takeover is never a single action.
The project initiation document from the China Intelligent Transportation Industry Alliance defines a takeover as the process where driving authority is transferred (partially or fully) from the autonomous driving system to the driver, while takeover performance is defined as the comprehensive evaluation of the driver's operational response during this process.
This indicates that a takeover involves a complete human-machine interaction chain encompassing system exit, request issuance, human perception, judgment, and operation. Its quality depends on whether the exit timing is reasonable, if the request is clear, and whether the driver was in a state capable of intervening at that moment.
Takeover performance is also not a unidimensional concept. It involves influencing factors, prediction models, improvement methods, and evaluation methodologies, requiring consideration of both subjective and objective indicators during assessment.
Yet in many tests, the takeover rate compresses this multidimensional process into a simple ratio: the number of manual takeovers within a specific mileage.
It does not record the system's confidence level at the time of takeover, nor whether the system had already attempted to handle the situation before exiting. It fails to distinguish whether the takeover occurred because the system objectively could not perform, or because the driver subjectively did not want to wait.
Zhou Guang, CEO of DeepRoute.ai, once described the significance of this magnitude difference in one sentence: "One takeover every few dozen kilometers versus one takeover every thousand kilometers are two completely different species."
But here too, "takeover" remains an undifferentiated, general concept.
If we do not break down what type of takeover occurs every few dozen kilometers versus every thousand kilometers, the two systems may simply not be comparable.
Furthermore, the industry currently lacks unified test scenarios and publicly available takeover data. Without standardized testing environments, open data, or objective safety scores, determining whether a system handles edge cases like sudden pedestrian appearances, cut-ins, or construction zones effectively relies entirely on automakers' self-reports.
Under these conditions, the takeover rate lacks an industry-wide quantitative passing threshold yet is used as a basis for ranking.
02. Why This Number Lacks Even Stability
The premise for the widespread use of the takeover rate is its perception as a comparable and reproducible objective metric.
But reality differs.
Takeover behavior heavily depends on individual driver characteristics. Drivers with lower trust in automated systems have a higher probability of actively taking over during critical operations; conversely, those with high trust may choose not to intervene even when the system approaches its capability limits. This means the takeover rate includes a significant proportion of driver-related variables unrelated to system capability.
The same system will show increased takeover counts if operated by someone more risk-sensitive, and decreased counts if operated by someone more willing to let go.
The statistical methodology remains unchanged, and the system capability remains constant, but the numbers fluctuate.
Currently, the takeover density for urban Navigate on Autopilot (NOA) remains relatively high. Statistics indicate that mainstream models require approximately 8 to 15 manual takeovers per 100 km, equivalent to one takeover every 7 to 12 km. Third-party field tests also show that the manual takeover rate for unprotected left turns approaches 50%, while right turns at intersections exceed 30%.
Given this baseline, fluctuations caused by driver behavior are sufficient to overshadow the limited differences between various systems. When rankings are based on the difference between 8 and 12 takeovers, few people question how much of that gap stems from the system versus the person behind the wheel.
03. How Exit Strategies and Inaction Minimize the Takeover Rate
The takeover rate only counts how often human intervention occurs, without recording what the system did prior to intervention. This gap leaves ample room for strategic adjustments. Since the metric only checks if a human took control, the system can manipulate whether it allows humans to take control.
One approach is proactive exit when confidence is insufficient—stopping operation and returning control to the human, which statistically adds only one takeover event.
If this exit occurs at the edge of a scenario where the human can easily resume control, it might not even be counted as a critical takeover.
From a system behavior perspective, this is boundary management—ceasing operation outside its own capabilities.
But from a metric perspective, proactive exit and system failure are lumped together in the same count.
Lack of system capability should ideally manifest as errors; however, once the system chooses to exit, it is no longer counted as an error. Instead, it appears as if the system recognized its limitations and prudently returned control to the human.
Another approach is for the system to neither exit nor execute actions. For example, failing to change lanes when necessary, refusing to merge, avoiding detours, and maintaining cruise control at speeds below traffic flow.
In such cases, the takeover rate definitely will not rise because the system avoids entering interactions that would trigger takeover risks.
These systems will inevitably look good on the takeover rate because they simply choose not to touch scenarios that inflate the number.
However, there is a vast difference between doing nothing and performing poorly.
Soochow Securities previously categorized takeovers during road tests into safety-critical and non-safety-critical types. The former refers to necessary interventions due to functional impossibility or imminent danger, while the latter refers to preventive interventions based on the driver's subjective judgment.
We will not elaborate further on this specific test report here; interested readers are encouraged to consult the original source.
04. Decision Quality Is Precisely What the Takeover Rate Misses
The takeover rate measures the frequency of human intervention, not the quality of the system's judgment prior to risk occurrence, nor does it reflect whether decisions made without triggering a takeover were correct.
Consider two systems: one has fewer takeovers but exits only at the last possible moment; the other has more takeovers, but each occurs early in the risk accumulation phase. The first wins on the takeover rate metric, while the second excels in decision quality.
A single count cannot capture this distinction, nor can it detect situations where no takeover was triggered, but the decision itself was wrong.
Ghost braking is the most typical example. The system does not return control to the human but executes an erroneous brake maneuver itself. Thus, it generates no takeover record but creates actual risk.
The takeover rate also fails to reflect what happens after a takeover.
One system might exit abruptly, but the timing is too late, leaving the driver an extremely short reaction window. Another system might request takeover earlier, providing sufficient time for human handling.
Both systems might have identical takeover counts, yet their actual risk levels differ significantly.
If the time window for a takeover request is insufficient to complete an effective handover, the gap between "a takeover occurred" and "a safe takeover occurred" is far more important than the difference in raw takeover counts.
05. What Is the User Feedback?
If a low takeover rate truly reflected system capability, systems with lower rates should be activated and used more frequently by users.
But this is not the case.
Data from the Ministry of Industry and Information Technology (MIIT) shows that since 2026, the penetration rate of L2 combined driving assistance features in domestic passenger cars has reached 70.5%, with NOA-equipped vehicles accounting for 34.2%.
However, a survey by Soochow Securities revealed that only 31% of owners frequently use urban NOA functions, 20% explicitly stated they dare not use it, and over 70% rarely or never activate it.
Zhang Xiaoyu, Executive Vice President of Changan Automobile, also noted at the TEDA Forum that while urban NOA penetration reached 20.9%, the usage rate was less than 6.3%.
Retention rates for highway NOA are similarly unsatisfactory. Chen Yichi, CTO of KPIT Technologies, pointed out that user retention after activating highway NOA ranges between 60% and 70%, meaning three to four out of ten users who tried it eventually gave up.
The discrepancy between installation rates and activation rates points to a question the takeover rate cannot answer: Does the system exist in the feature list, but are users willing to let it work in real-world scenarios?
A system with very few takeovers but a high proportion of users manually disabling it likely reflects not superior capability, but rather users making their own judgment through their actions.
The fact that the person behind the wheel chooses not to enable it speaks volumes about the system's usability in real-world conditions.
06. Separating Safe Takeovers from Total Takeovers
The takeover rate is not valueless. The main issue lies in indiscriminately counting interventions of different natures under a single number, then using that number to answer questions it cannot address.
MPCI (Miles Per Critical Intervention), proposed by Yu Kai of Horizon Robotics, incorporates only takeovers related to collision risk, excluding all other manual interventions.
Within this framework, top-tier urban NOA systems in the industry have just reached an MPCI of 100 km, corresponding to "usable" status; 1,000 km corresponds to "good"; achieving "loved by users" and adapting to Robotaxi-level L4 scenarios requires an MPCI exceeding 100,000 km. Completing this series of magnitude leaps is estimated to take 4 to 5 years.
Compared to focusing solely on the takeover rate, regulatory safety compliance requirements hold greater significance.
The "Safety Requirements for Combined Driving Assistance Systems in Intelligent Connected Vehicles" (GB 47955—2026), effective January 1, 2027, imposes mandatory requirements on functional safety, test coverage, OTA management, and Driver Monitoring System (DMS) surveillance.
Once compliance becomes a common constraint for all players, the true differentiators will be those who incorporate decision quality, traffic efficiency, and actual user activation willingness into their evaluation systems.
The takeover rate can continue to exist, but it is better suited as a clue that needs to be deconstructed and cross-validated, rather than serving as the goal for R&D direction.