Tesla Robotaxi rams through bollards weeks after ‘impeccable’ claim

A Tesla “Robotaxi” running with no safety monitor onboard drove straight through a row of bollards protecting a closed-off lane in Austin, then kept driving.

The video surfaced Monday, less than a month after Tesla’s head of AI told investors the Robotaxi program has an “impeccable safety record” with “zero notable incidents.”

The failure is upstream of perception. Tesla doesn’t use the detailed prior maps Waymo relies on, so the car works out the road layout on the fly. It read a right-turn lane, committed to it, and then couldn’t work out how to reverse back out of its own mistake.

That’s the part worth sitting with. Faced with an obstacle it had already detected and stopped for, the car’s answer was to drive through it.

For scale, Waymo has logged more than 220 million rider-only miles and is running roughly 500,000 paid trips a week. Tesla’s entire cumulative unsupervised total is less than 0.2% of Waymo’s.

And the Tesla fleet doing that driving is tiny. We reported in June that only 14 vehicles were operating unsupervised across all markets, down from a peak of 25. A clean record over that little exposure isn’t a safety result. It’s a sample size.

The car saw the obstacle. It stopped for the obstacle. It sat there. And then it decided the right move was to go through. You don’t fix that with a better sensor, because sensing wasn’t the problem.

Now run the same logic against concrete bollards. Or a barricade at a street festival. Or a work zone with people behind the posts. The car can’t tell plastic from concrete.

Jesus. TSLA Robotaxis are rolling deathboxes. No wonder Nevada only permitted TSLA for 10 Robotaxis total.

1 Like

Actually you do. With Lidar, it would override anything else because it finds “a solid object” and overrules. Whatever the issue was with “vision only”, the ultimate rule would be to “discuss” between sensors and let either one of them decide “solid object, stop immediately.”

That’s how multi-sensors work. There are certain absolutes, and if one sensor fails for whatever reason, the other would take control. It’s just two linear gates, if either one says ‘no go” then you don’t go.

It’s the basis of the thermistors in your phone, any smoke detector labeled “multi-sensor”, or other things safety related. (I use those two because people get jonesed when I talk about aircraft or refineries or oil platforms or water systems or factory machinery or chemicals plants or other higher cost applications.)

Heck, a newer washing machine uses multiple redundant sensors to check water levels with electronic backup level/overflow safety switch as well as pressure and sometimes even visual observation to prevent floods if the primary tube clogs.

But I’m sure it can’t be done, because … what, again?