Tesla Robotaxi Deployment in Austin

What I'm thinking about safety as the date approaches

Here are my thoughts on safety considerations for the impending Tesla robotaxi deployment in Austin. As I write this Tesla’s target date is later this week, on June 12.
(Edit: Now it is June 22. Perhaps it will change again, but the below observations probably won’t regardless of which day on June it launches.)

Source: https://digitalassets.tesla.com/tesla-contents/image/upload/IR/IR/TSLA-Q1-2025-Update.pdf

The reality is nobody outside Tesla can know whether the robotaxis will be safe at the time of launch. Moreover, things Tesla is saying leave a lot of room for interpretation, so everyone is just guessing at this point. Even NHTSA has a lot of good questions to ask, with Tesla having picked a launch date before their response is even due.

Tesla may have already launched. From a safety point of view, the actual launch happens the first time Tesla removes the in-vehicle safety driver. There is no such thing as driver-out “testing” when it comes to safety. If there is no driver in the vehicle, then any driving mistake can cause a crash, regardless of whether the company calls it “testing” or not. The rest is a matter of scale and narrative.

Apparently there have been some operations with a person in the passenger seat, but with an unclear ability to control the vehicle. At some point that person comes out, and that is the point at which other road users are sharing the road with a driver-out robotaxi, regardless of whether it is labeled “testing” or “operation”. It would not surprise me if this has already happened, but we won’t know unless Tesla tells us.

Remote Assistants. The big question for safety is the role of remote assistants in this deployment. There are several important pieces to this:

• Does the remote assistant actually drive the vehicle sometimes? For example, do they have a remote steering wheel to operate the vehicle some of the time? Companies such as Waymo make a big point that their remote assistants do not have a steering wheel. If Tesla remote assistants have the ability to directly control steering (via steering wheel or otherwise) that makes them remote drivers. This does not mean Tesla won’t have a robotaxi. Chinese robotaxis have had remote drivers with steering wheels as well. But from a safety point of view there are very serious concerns about communication lag and dropout that must be addressed. Eventually there will be a communication drop-out or high lag at the wrong time, presenting a hazard that could result in a crash. The question is what Tesla has done to mitigate this risk. Saying, for example, that there are two mobile data carriers operating in parallel is only a start, and not sufficient for a number of reasons. Assuming they have done that.

• Is the remote assistant expected to continuously monitor the vehicle at all times? If the computer driver asks for help and the remote operator does not need to pay attention most of the time, that is on a par with what other robotaxi companies in the US have designed. If, however, the remote assistant needs to notice something is going wrong and jump in to intervene when the robotaxi does not identify a problem occurring, that is a supervised vehicle rather than an autonomous one.

• If the remote supervisor needs to intervene, is it a remote test driver for a robotaxi prototype (if Tesla says this is a Level 4 feature)? Or a remotely supervised Level 2 vehicle automation feature? This will make a big regulatory difference. For human-driven vehicles Tesla has tried to have it both ways between regulators and stockholders.

• Where are the remote assistants located? If they are outside Texas, there might be jurisdictional issues if there is a crash or road rule violation. If they are outside the US, things get even more tricky here. I’d expect the remote assistants to be in Austin at first to minimize control lag. But later on as operations scale this question will become more important.

• Are there plans to (eventually) assign multiple vehicles to each remote assistant? If so there are issues regarding maintaining situational awareness for multiple vehicles and what to do if multiple vehicles need assistance at the same time.

What does safe enough mean for Tesla? If Tesla says the vehicles are safe, what exactly do they mean by that? As we found out in San Francisco with other robotaxis, safety is a complex and nuanced thing. If Tesla wants us to trust that they are safe, they need to bring some serious metrics to the table rather than handwaving.

I am much more worried about the safety of vulnerable road users (pedestrians, scooter riders, etc.) than of passengers. Passengers are protected by all sorts of crash safety technology. Vulnerable road users, not so much. Even miles with nobody in the robotaxi puts other road users at some risk.

When will the first crash happen? This one nobody can know. If Tesla only deploys 10 vehicles, they can operate without a crash for a long time even if their technology is quite poor on safety compared to a human driver. At ballpark 100 million miles per fatality in the US, ten vehicles simply cannot drive enough miles to have reached the first expected fatality. As a practical matter, we should see no loss events at all for a fleet of 10 or 20 vehicles this year, even if they are just average compared to human drivers. But edge cases have a way of sneaking up on companies, so we’ll see how it turns out for Tesla. Loss events should only be expected due to increasing exposure to hazards as the technology scales, not with a small pilot deployment.

For everyone’s sake I hope we do not see an injury involving a Tesla robotaxi this year. Certainly if we do that will signal a big safety problem for Tesla if it happens when they only have a few tens of vehicles on the road.

Other topics: There are a lot more speculative topics that come up, and the answers for all of them are: we really don’t know. But here are some questions that might be interesting to seek answers for: Does improved performance come from end-to-end machine learning for robotaxi service that has a high content of Austin-specific training data, and is it the same training data for retail customer FSD? What is the maximum operating speed (below 20 mph helps with safety)? What are the go/no-go criteria for release to ensure public safety? What are the criteria for declaring a stand-down in response to a crash? Does Tesla accept that it has a duty of care for the safety of road users with regard to tort law? Does Tesla claim this is a Level 2 vehicle or a Level 4 vehicle?

For those who like keeping score, here is the prediction I made a while back about what we will see deployed.

For those who want to dig into the nuances of remote driving, I have a piece about that.

I may update this piece as more information becomes available or I run into a question not covered above. If you are a journalist and can’t get in touch with me by deadline, please feel free to quote from the above material.