I’ve been saying for a while that LiDAR and radar have weaknesses too, and that they all have real downsides in adverse weather. LiDAR is literally active light; rain, fog, and snow can scatter and attenuate it. And radar isn’t “magic vision” either: automotive radar is great at range and relative speed, but it can struggle with angular resolution, clutter, and confidently classifying what an object is, especially in messy weather. Even big weather-radar systems don’t produce a perfect, high-resolution picture at long range; detail falls off the farther out you look.
In reality, I hope Tesla can get to unsupervised FSD on cameras, but I don’t know enough about ML to say where the real bottlenecks are. What I can say is the current camera suite still seems constrained by basic sensor limitations (resolution, dynamic range, low light). If they ever add anything, I could see higher-res/better optics helping at a distance and a forward-facing IR/thermal-style camera helping with certain night/contrast edge cases. Not saying that “solves” autonomy by itself, just that it could reduce some of the situations that are inherently hard for visible-light cameras.
And on occlusion: this is the part people hand-wave. Cameras get blinded by spray, mud, bugs, condensation, and glare—the same stuff that blinds your eyes. You can’t “AI” your way through a lens that’s physically blocked. The real fixes are hardware: overlapping viewpoints so one dirty camera doesn’t sink the stack, plus active camera hygiene washers/cleaning jets, lens heating/defogging, coatings, etc.
What’s telling is Tesla clearly seems to treat this as a real engineering requirement on purpose-built platforms: we’ve already seen them emphasize additional sensor-cleaning/robustness measures on their robotaxi-focused vehicles (and the Cybercab concept). That’s basically an admission that “vision-first” isn’t just neural nets; it also needs the equivalent of windshield wipers for the car’s “eyes.”
www.notateslaapp.com
