By "robotaxi" I think they meant loosely "personal self driving car," not an automated taxi service.

Waymo's overstated[1] success has let self-driving advocates do an especially pernicious bit of goalpost-shifting. I have been a self-driving skeptic since 2010, but if you had told me in 2010 that in 10-15 years we have robotaxis that were closely overseen by remote operators who can fill in the gaps I would have thought that was much more plausible than fully autonomous vehicles. And the human operators are truly critical, even more so than a skeptic like me assumed: https://www.nytimes.com/interactive/2024/09/03/technology/zo... (sadly the interactive is necessary here and archives don't work, this is a gift link)

I still think fully autonomous vehicles on standard roads is 50+ years out. The argument was always that ~95% of driving is addressable by deep learning but the remaining ~5% involves difficult problem-solving that cannot be solved by data because the data does not exist. It will require human oversight or an AI architecture which is capable of deterministic reasoning (not transformers), say at least at the level of a lizard. Since we have no clue how to make an AI as smart as a lizard, that 5% problem remains utterly intractable.

[1] I have complained for years that Waymo's statisticians are comparing their cars to all human drivers when they should be comparing it to lawful human drivers whose vehicles are well-maintained. Tesla FSD proves that self-driving companies will respond to consumer demand for vehicles that speed and run red lights.

Waymo is at the threshold of how speech recognition reached a tipping point through decades of grinding effort. I worked on speech recognition applications in the '80s when the best of it was just barely usable for carefully crafted use cases. Now taxis can be automated well enough if you have the right sensor suite and an exquisitely mapped environment. Waymo can do this for 0.05% of ride hailing riders, if 28M Ubers per day and 100k Waymos per week are reasonably accurate. This is like when connected speech started working with a good mic in a quiet environment. Now automated multi-speaker transcription with background noise, crosstalk, and whatever mic you got works. Back in the day some serious people were convinced speech recognition on that level would be impossible without solving AGI.