Comment by simondotau | Hacker News Reader

Comment by simondotau | Hacker News Reader

Parent Full thread

I didn't say FSD beta was "ready". I merely assert that it is demonstrating remarkable competence in spatial sensing. If spatial sensing isn't a problem—and clearly it is not—it makes no sense to add LiDAR, which only does point cloud spatial sensing. The real "hard problem" of self-driving is in the path planner, not spatial sensing.

I hadn't heard of AutoX. I watched a couple of their videos. Despite being videos cherry-picked by the developer they show a car being driven in cities at slow speed and on faster arterials seemingly unable to drive without ping-ponging. No evidence that these cars could ever drive on roads they haven't been directly trained on.

By comparison, thousands of videos of FSD beta are being posted by hundreds of real world vehicle owners driving intentionally challenging routes in dozens of distinct geographic jurisdictions. It is showing remarkable progress, but of course there's still limitations; understandable when exposed to such a massive long tail of edge cases.

E.g. https://www.youtube.com/watch?v=0g4TnvvuGqM

The current SOTA in monocular “spatial sensing” (3D object detection) is just much less reliable compared to LiDAR-only systems.

You can’t conclude that spatial sensing isn’t a problem because it didn’t crash with a monocular camera. The test is whether it accurately detects all relevant actors/objects in the scene and does so reliably.

From a couple FSD videos I’ve seen, the issues center around perception (failing to get the right lane lines, certain detections disappearing briefly, etc.)

> You can’t conclude that spatial sensing isn’t a problem because it didn’t crash with a monocular camera. The test is whether it accurately detects all relevant actors/objects in the scene and does so reliably.

Indeed, which is why Tesla have put a lot of effort into showing a visual representation of its object detection. Unlike many (most) other players in this space, Tesla isn't afraid to let people see what their system is capable of in literally any situation their customers wish to drive to.

Most of their competitors are hiding behind L4 which, because of how machine learning works, makes systems appear better than they are. Overfitting is a great way to make rapid gains, but much of it is illusory.

> From a couple FSD videos I’ve seen, the issues center around perception (failing to get the right lane lines, certain detections disappearing briefly, etc.)

Even if true, LiDAR wouldn't help as it cannot see painted lane lines and is extremely weak at object identification.

Looking at the most recent FSD beta (10.12+) as shown on YouTube by real customers, it's quite clear that detection is already damn close to exceptional. When FSD does fail, it's usually either out of an abundance of caution (it is extremely cautious around pedestrians) or because the path planner wasn't making a good decision. It's almost never because the car has failed to see something that LiDAR might have.

> Overfitting is a great way to make rapid gains, but much of it is illusory.

I'm not sure why you think this is the case for L4. Do you mean since it can be geo/weather restricted the learning task is easier? Surely weather plays a large role but I don't see why other L4 companies necessarily overfit more (also how does overfitting provide rapid gains?). If you overfit when training for your L4 system...it's not an L4 system.

> Even if true, LiDAR wouldn't help as it cannot see painted lane lines and is extremely weak at object identification.

True, but:

A) Again, the test is reliable detection which is simply out of reach atm from single-sensor input

B) HD maps help a lot with this issue since the lane lines become less important

When you limit your service area, it becomes feasible to train your models on pretty much (or literally) every road in the service area. Every single permutation of weird intersection it will ever come across. If your L4 car only ever has to drive in San Francisco, it doesn't matter that this ML model would have no hope when introduced to Salt Lake City. The model never has to spend a moment excluding any of the millions of potential things it has never been trained on and will never see.

Both the ease of data collection and the absence of contextual noise makes L4 machine learning an order of magnitude easier.

Can you point me towards any recent FSD Beta video where there was an important failure of object detection?

Here's a list of robotaxi companies https://en.wikipedia.org/wiki/Robotaxi#Testing_and_revenue_s..., some of them operational, Level 4.

I would be very surprised of any of them use AI vision without Lidars. That's what my point is about, Tesla is lagging behind. Wake me up when it gets Level 4 certification.

The SAE levels don't measure technical sophistication. In fact when it comes to machine learning, an L4 system is substantially easier to implement than L3 or L2 because the task is monumentally less difficult and you can overfit your way to perceived success.

Whether you (or I, or anyone else) thinks Tesla is lagging behind or not has no bearing on the actual reality, and it won't change whether or not Tesla's approach is ultimately successful. Hopefully we're both fortunate to live a few more years and see what happens first hand.

Reply on news.ycombinator.com