Hacker News Reader

> How many thousands of miles of snow driving data does Tesla possess, compared to the #2 player?

It depends. If they're not recording the right things, they have the same 0 bytes of data as everyone else.

There's a not insignificant chance that their sensor data is insufficient to perform the tasks, and the recorded data won't be usable for training a system with whatever sensor is needed; in which case, no advantage for logging customer movements for N years.

Yep, Tesla is placing all their eggs into one basket with purely visual self-driving.

Visual-only also delivers less fuel for lawyers to armchair quarterback, and that advantage applies even in the limit of good AI.

Man, those humans are so dumb thinking they can drive safely just by looking at the road /s

Let me know when Tesla's cameras can put their arms up to shield their eyes from glare. Or move over a few inches to see around a big truck. Or react to a stopped big red fire truck with flashing lights. /s

Why does Tesla bother with radar and ultrasonic sensors? Humans can drive without them.

I just find it strange that Tesla is criticized for relying on visual self-driving systems when most humans rely primarily on vision to navigate.

Well but Tesla cars don't have brains. Maybe the lack of lidars makes the task 100 times harder, i.e. de facto impossible. We don't know. This is a big risk for Tesla but they didn't have a choice. If lidars were as cheap as cameras, of course they would use them.

On the other hand, Waymo et. al. are not risking anything by using lidars. If they turn out to be redundant, it should be easy to remove them cheaply.

This isn’t true: if you build a system for highly sophisticated inputs, simplifying the inputs can break all sorts of assumptions about the accuracy of the input data and other such invariants. If you start with a more primitive set of sensors and go as far as possible, adding more accurate sensors seems less risky.

Maybe in general, but I don't think you're right in the case of self-driving systems. Waymo says that perception is the easy part and most of the work is elsewhere. So in the absolute worst case scenario, they would have to develop the perception part from scratch. At Tesla, this took 2 years.

BTW, all companies that can, use lidars. Even MobileEye says that their L4 system will use lidars and they've been working on a camera based system for 2 decades. Tesla doesn't use lidars because they can't.

I mean - I have no idea what’s going on at Tesla, but its very likely they’re doing some internal work with Lidar. They surey understand the risk with w vision-only approach. Elon, and Karpathy, and Laettner previously, aren’t fools. The problem for Tesla is they can’t be too Lidar dependent because nobody mass-produces them

Tesla can't use lidars because: 1. It would make the cars much more expensive. 2. They would have to refund all people who've bough FSD or provide them a free upgrade.

These are fairly artificial constraints.

1. Lidar is expensive for everyone. It’s not like Tesla will pay more than other manufacturers. They’d prefer not to use Lidar, but if cameras just don’t cut it, I don't think they would die on that hill.

2. I’m sure there is some fine print around the current FSD being for camera based sensors only. They can easily say the car tops off at L4 for cameras, and L5 needs Lidar, if we come to that bridge

Yeah, it's expensive for everyone but the point was that the FSD product is de facto impossible with lidars. Therefore, they have to pretend that it can be done with cameras.

Because engineering is about trade offs.

Why isn't every car body made from carbon fiber? It's lighter and sturdier that alternatives. It's also more expensive so you can put it in $60k car but not in $25k car.

Radar and ultrasonic sensors are cheap, low power and plentiful.

Putting several lidars on a car would be too expensive, would use too much power (less range) and there isn't a supply chain to make 2 million a year just for Tesla (500k cars x 4 lidars per car).

It's just isn't possible to put a lidar today on a mass market car and Tesla wants to solve self-driving before lidar becomes cheap enough.

Fleet data of Tesla's scale probably doesn't give you a substantial advantage. MobileEye seems to be ahead of Tesla with a lidar-less system without such data (but they have partnerships with car makers to collect mapping data). From what I hear, companies like Waymo don't want more data, the bottlenecks are engineering.

In self-driving, you basically have 3 parts - perception, prediction and planning. Perception is the easy part (if you have lidars). Tesla and AFAIK others rely heavily on NNs for perception. The real challenge is prediction and planning. To what extent are NNs involved here in Tesla's system is unknown, probably not that much. Overall, I don't think that Tesla has a fundamentally different approach than others.

Hasn’t Waymo already won? They’re driverless now.

No, they still use following cars and remote assistance

But they're driverless. This means a very high confidence in the system. Hopefully, they'll reach total confidence next year and get rid of chase cars.

> Tesla’s approach of vision based, AP disengagement-led training will scale better ... [about weather] they’re special cases for other companies (from what’s been revealed; Comma might be an exception).

> all it takes for Tesla’s NN approach to pay off is one leap in explainability.

It's not vision and disengagement-led training that's conferring the scale advantages you're talking about, because other AV companies actually do train on and attempt to drive in various kinds of weather, and safety drivers do the same thing normal drivers do except that they're more trained. The difference is the approach Tesla has to safety (vs the others). Which is to say: Tesla doesn't worry about training their supervising drivers, at the risk of those folks' lives.

Tesla also doesn't seem to worry about _proving_ the safety of their systems: explainability isn't simply going to be a matter of a new algorithmic breakthrough. How do you prove that your software is safe if you don't sample heavily from different operating domains you want your car to work in, and then either simulate on those miles or actually test your car? Before letting a driver behind the wheel of an autopilot, companies like Cruise/Waymo/Aurora all do manual driving first in a new domain, and then move to safety drivers (even 2 to start, for extra safety). This is not because their software is incapable of doing the same things as Tesla, but because it's very worthwhile to sample disengage rates and assess safety criticality to various issues before allowing the car to act autonomously. And even then there are safety nets in place to further mitigate risk. (You're probably right that L5 is a long ways off, but this is how we'll get to L4.)

Of course it's more "scalable" to avoid all this and put the responsibility of risk on users who, as your system improves, become less and less attentive. And so this is what Tesla does. I suspect solving the long tail of issues is going to be more of a challenge for Tesla because at least in my experience it's easier to {convert a naive rules/simple model --> complex model} than to {debug the complex model --> handle failure modes with logic}.

Very valid points. I think the key is that you’re going to need a complex model for full self driving. No way around that - even if Waymo starts with a rule based model, they’re going to have to work with black boxes and debug them at some point (and they do).

Tesla just decided to jump into the deep end from the beginning.