Which is certainly better than just having a human remote drive.
But it’s still not the impression they’ve been giving. It’s been an impression of full automation (ignoring getting stuck) and if it’s not navigating on its own that’s disingenuous.
This approach has two benefits: it can be unstuck without sending out a physical driver and while collecting training data, and it efficiently lets m humans control n cars with a wide range of acceptable m and n values.
It's intended for the ratio of m:n to smoothly shrink as the software gets better, but m will always be greater than zero.
* Unless it gets super stuck, then a human drives out and gets into the physical driver seat and takes over