Hacker News Reader

Comment by Animats | Hacker News Reader

The fundamental problem with Tesla's driver assistance system is straightforward - it won't brake for anything other than a car rear end. So far, Teslas on autopilot have hit at speed at least two semitrailers, a fire truck, a street sweeper, and various other objects partly obstructing a lane. The NTSB has pictures of most of these.[1] In most cases there's no braking at all - the thing just plows into the obstacle at full speed.

Musk's decision not to use LIDAR is paid for in blood.

[1] https://www.ntsb.gov/news/events/Pages/2020-HWY18FH011-BMG.a...

View on HN

I'm fairly critical of Tesla in general because they're selling snake oil, with the full self driving promise and whatnots, but this particular critique targeted uniquely at tesla seems misguided since every other System of highway assist whether with LiDAR or Doppler radar ignores stationary objects

Stack several technologies together instead of relying on just one. Edge cases will be more easily identifiable because one technology will have high confidence about an obstacle that another does not perceive.

that works up to a point.

if lidar and radar don't detect an object, but stereo vision do, what do? you use consensus, or you go with at least one sensor.

there's a lot of tradeoffs involved, including stacking false positives.

EyeSight (Subaru's assistance system) uses stereoscopic cameras, so it can handle large stationary objects.

LiDAR can distinguish between on and off-road objects, so it doesn't have to ignore stationary objects. I don't know of a LiDAR-based highway assist that's commercially available, though.

Finally, because we're a bunch of pedants here, radar-based systems will register stationary objects that were formerly moving. This is how "follow to stop" systems know when to stop. :)

I was wondering, as I use EyeSight on my Outback and it seems to detect plenty of stationary and odd shaped objects (like an old, empty flatbed semi with broken tail lights) which I didn't expect it to.

> Musk's decision not to use LIDAR is paid for in blood.

LIDAR is not the only way to record depth information. Saying that LIDAR is some kind of savior device is heavily into "solve all your fusion problems with this one weird trick" territory.

Okay, but from the statement, whatever alternative they're using isn't working

Don't be ridiculous. It's more like "see better with different glasses."

Do you have a paper comparing states of the art flash LIDAR to real time high density stereo that makes you think it's better? Because my experience with the field has been that lasers aren't better for this except for very specific non-real-world scenarios.

Why is a paper required? Many of Tesla's early self-driving accidents involved low contrast scenes where the system was unable to distinguish, e.g., a white object from a wall, or a blue object from the sky. This is always a weakness with optical, even for our own eyeballs.

LIDAR is not the only way to record depth information

When you mix two or more different ways of measuring depth you can increase the reliability of your system dramatically. See "sensor fusion". Optical-only has weaknesses. LIDAR-only has weaknesses. Optical plus LIDAR would be far more robust than either alone.

Reply on news.ycombinator.com