Why Full Self-Driving Still Has a Long Way to Go (FSD 13.2.9)
I recently took a short 2-mile drive to a restaurant from my home, and despite the simplicity of the route, I had to take over the wheel four separate times due to safety-critical issues. Here’s what happened:
1.Missed Speed Bump at 30 mph:
FSD drove straight over a large, clearly marked speed bump. It was marked with reflective road stripes, vertical reflector posts, and an oversized warning sign. Despite all of this, the system failed to recognize it. Not only is this enough to damage tires or suspension, but none of it was even visualized. You’d spill your coffee, or worse, if FSD were in control.
2.Failure to Stop at a Standard Stop Sign:
It completely ran through a stop sign at a pedestrian crossing. This wasn’t an ambiguous or unusual sign. It was a standard red octagon with white lettering, identical to others throughout the region.
3.Incorrect Lane Selection at a Marked Intersection:
At a three-lane intersection (left, straight, right clearly marked), FSD drove down the middle lane and attempted a left turn. It was fortunate no oncoming traffic was present. This could have caused a side-impact crash.
4.Entered a Clearly Signed “Do Not Enter” Street:
It turned into a street marked with a standard “Do Not Enter” sign and simply stopped mid-road, unable to proceed or recover. Again, no oncoming cars, but there could have been.
5.Failure to Exit Driveway:
Before any of this, it took four failed attempts to perform a simple three-point turn just to get out of my driveway. It eventually gave up. I didn’t count this among the four safety interventions above since it wasn’t dangerous, just frustrating.
For context, I’m using FSD Beta 13.2.9, and I’ve posted video evidence of some of these issues previously with version 13.2.8.
⸻
The Bigger Picture
These experiences underscore a key reality. We are still far from achieving true unsupervised driving. Tesla’s Full Self-Driving handles structured, high-traffic, well-mapped areas far better than it does local, less-structured environments like neighborhood streets, driveways, parking lots, and suburban intersections. The reason is that the density of other vehicles and pedestrians provides crucial context that boosts performance via inference. In contrast, the system struggles where it must rely more heavily on road rules and signage alone.
So while some FSD videos show flawless drives through cities, they’re often cherry-picked from high-traffic, geo-trained zones. My drive, in contrast, demonstrates how easily FSD can unravel when the environmental scaffolding disappears.
I remain optimistic and bullish long-term. These problems are solvable. But Tesla’s own missed timelines, like the undelivered 3x context window increase promised over six months ago, suggest serious ongoing challenges. It also hints that training resources are being funneled into a limited set of geo-fenced areas to get Robotaxi viable in specific launch cities (for example, Austin), rather than generalizing the solution for nationwide rollout.
If Robotaxi Austin can navigate freely into neighborhoods, plazas, parking lots, and private driveways without geo-fencing or restrictions, that will be a true milestone. Until then, let’s not pretend we’re already there.