Hacker News Reader

Comment by njarboe | Hacker News Reader

Mandating how a certain outcome should be achieved generally is not the best (efficient, quickest, cheapest, etc.) way in getting to that outcome. Autonomous certification should have its safety requirements for each level and require the vehicles meet that standard.

Autonomous vehicles have the potential to reduce deaths and injuries dramatically and increasing the time to get there will cause thousands to millions of more deaths and an order of magnitude more injuries, not to mention property damage.

View on HN

I don't know. In general I think you're going to need some balance between requiring certain outcomes and requiring certain specific implementation details. It might be possible to make an automobile out of sufficient thick and squishy parts that it could pass reasonable crash safety tests without having a seat belt, but it's still not ridiculous for regulations to require that automobiles have seat belts.

(I have no idea where lidar for self-driving cars falls within that balance, although I can't help but notice that human drivers do not utilize lidar.)

Well, lidar is just a stand-in for the depth perception humans already do. It is absolutely valid for regulators to require that the systems on an autonomous car do depth perception, and provide a standard to prove that the system works well.

I suspect that right now, ML based approaches to depth perception would not pass the test to standards we would like, in practice requiring lidar. But, this would also provide an option to switch to an ML based gets good enough.

This is a common misperception. You cannot estimate the shape of simple things in low light. Even humans are susceptible to optical illusions and are effectively blind during certain times of day (try driving in the direction of the sun during sun rise/set). Lidar is not susceptible to any of these things. Similarly ML will never solve low light or direct light situations, that is a physical limitation of cameras.

There are popular cars in the US that use two cameras for their self-driving functionality. Kinda like humans.

Those cameras are fixed whereas eyeballs have multiple degrees of movement via the eye socket and skull mounting apparatus allowing very sensitive parallax measurements. It's really not the same at all.

Autonomous vehicles do of course have the potential to reduce deaths but there are lots of examples of this being the preferred method of regulation in the auto industry. For example if you want to sell a motorcycle in the US it has to have a kill switch in a specific place on the right handlebar. You don't (as a manufacturer) get to innovate about this because it's a safety feature and a first responder needs to be able to locate the kill switch so it always has to be in the same exact place so they can always find it. If you want to sell a car in Europe (I believe) it has to have seatbelts, ABS and a catalytic converter that meets a particular specification etc etc.

This isn't necessarily that strict of a regulation. They could require the sensor be on board but not actually regulate any of its usage. It could be a scenario like the Model 3's rear view mirror camera, where it wasn't used at all for the first two or so years and then later got a software update that started interacting with it. The important part is that the LIDAR unit is physically wired in to the system so that they can start using it in software whenever they feel like it's necessary.

Reply on news.ycombinator.com