Hacker News Reader

Comment by uptownfunk | Hacker News Reader

I think Tesla is highly underrated in this space. All Tesla cars now have sonar and radar capability built in, in addition to the extensive camera system.

As someone who frequently commutes in Teslas, I would say Tesla is closest to actually deploying something close to self-driving to the general population out of all the current self-driving competitors. Furthermore, I imagine they're collecting data on all the Tesla cars currently being driven.

I have seen the model updates substantially improve the auto-pilot capabilities with every update. And as someone who works with lots of quantitative data on a daily basis, I think there is a fundamental difference between training on billions of miles of simulated data vs. actual data. Simulated data may reduce overfitting, but there is still scope to overfit compared to raw data gathered from the field. Simply put, you won't generalize as well to data that's been simulated vs. field-generated data.

View on HN

You can train on a hundred billion miles of real world data - you'll still be wasting your time if you need LIDAR and you don't have it.

> you'll still be wasting your time if you need LIDAR

This is not for certain. I'm quoting Tesla's CEO, but it's suggested that LIDAR leads to optimizing for local minima. It means that LIDAR can get you 90% of the way there, but there's still going to be a lot of situations where LIDAR might not be capable.

The problem that needs to be solved is vision. Humans only have vision, and they can drive with it. Radar and LIDAR are nice-to-haves, but until you solve vision, you haven't solved self driving.

> The problem that needs to be solved is vision.

"Solving vision" is a 10x more harder problem than building a practical self-driving car with LIDAR. We might get it working with LIDAR in 2-5 years, but without LIDAR in 20-50 years.

Arguably Alphabet (Google + Waymo + Deepmind + co) is the largest group of AI and ML experts in the world. If someone is going to figure out general computer vision, then it's likely to be them - and if they are working on LIDAR first then there must be a good reason why.

I believe Tesla is promoting LIDAR-less systems just because it's expensive bulky hardware, and they can't possibly sell cars with it at a reasonable cost. Calling cameras and radars "full self-driving hardware" is a good marketing tactic, but without any guarantees that they can build the software for it in the next 20 years.

The issue from my (admittedly little) experience is that it's all about edge cases and reliability.

It's easy to make a demo with cameras only but to deal with the countless unimaginable edge cases, you need LIDAR.

The other factor is that you need damn near six sigma accuracy for ML algos for robust SDCs, and while it's theoretically possible to match LIDAR with cameras and current DL, it's impossible to do so with the needed 99.99 percent accuracy for autonomy.

> I'm quoting Tesla's CEO

Ah yes, the famously practical and emotionally stable CEO.

You might be interested to know that the last Tesla crash in California was pretty much directly caused by lack of LIDAR

I'm not saying you do or don't need LIDAR.

I'm saying that if you _do_ need LIDAR, a training data set without it ain't worth much, no matter how large it may be.

No they aren't. Teslas only self drive on the highway, which is great, but wouldn't last a minute in a more urban scenario with crosswalks, lights, tight merges, etc. And radar is not good for these scenarios either, because of the huge amount of false positives with the system.

Reply on news.ycombinator.com