You can't measure this, because it has drivers behind the wheel. Even if it did three "pedestrian-killing" mistakes every 10 miles, chances are the driver will catch every mistake per 10000 miles and not let it crash.
But on the other hand, if we were to measure every time the driver takes over the number would be artificially high - because we can't predict the future and drivers are likely to be overcautious and take over even in circumstances that would have turned out OK.
The only way to do this IMO is by
- measuring every driver intervention
- only letting it be driverless and marketable as self-driving when it achieves a very low number of interventions ( < 1 per 10000 miles?)
- in the meantime, market it as "driver assist" and have the responsibility fall into the driver, and treat it like the "somewhat advanced" cruise control that it is.