Tesla’s Current Status From an Engineer’s Perspective
I’m currently working as a junior automotive engineer. From that perspective, I want to tell you how far Tesla has come right now.
The Gap Between Hardware and Software
First, as you know, the normalization of autonomous driving is already set in stone. The hardware performance still hasn’t caught up, but compared to that, software performance has improved a lot. As you know, it was initially divided into LiDAR configuration and radar configuration, but practically speaking, Tesla is also learning from data including LiDAR and actually using it, developing as a technology that synthesizes all sensors. This is an established fact.
Deep Learning-Based End-to-End Architecture
But with the rapid development of deep learning, End-to-End technology has advanced, and Tesla has rewritten the code from scratch starting from version 13, so at this point, highly advanced autonomous driving has been implemented. If I had to point out a major critical weakness, it’s deep learning development—the fact that you can’t know what the basis is for judgments made based on deep learning.
Autonomous Driving With Extremely Low Human Error
Nonetheless, it’s a fact that autonomous driving has been implemented at a level where human error is practically extremely low, and I’m going to look at that in detail.
Human Neurons vs Computer Neural Networks
We humans are growing somewhat through neurons in our brains. Growing means analyzing external facts, external objects, and all moving things on roads to make judgments. But computers lack the computational capacity and logical structure to think like this, so they have to judge with only 0s and 1s. But by creating a neural network—a deep, multilayered structure—we can implement the brain’s way of thinking in the same way. That can be said to be the beginning of ImageNet starting in 2013, that is, deep learning neural networks.
Recognizing 3D Space From 2D Images
But how can you simply judge these facts? Let’s think about an image for a moment. Let’s say we’re taking an image from the perspective of a driver. There’s something rotating in front and there are objects. The data you need to extract here is how far away the object in front is—the absolute coordinate value. If you’re trying to figure that out, one image is enough. Why? Because with a single image, you can’t extract that depth—2D or 3D data. That’s why through radar and then cameras, you analyze those images from multiple angles and estimate depth. This technology is called sensor fusion.
Computational Power and Hardware Importance
For this technology to advance at a high level, it requires enormous computational power, and that’s being concentrated in Tesla’s hardware, and China is putting the same kind of hardware in cars the same way.
The Absolute Advantage of Data
What’s important here—what’s most important—as you know, is data. Real driving data and data that’s been organized to a certain degree, data that’s been consistently accumulated in a specific format. In terms of quantity and quality of that data, they’re overwhelming it, and it’s practically impossible to catch up.
New Frameworks and Explainable AI
To break through, a new framework has been proposed. They’re going to get data from virtual worlds, and they’re going to put out explanations every time they make a judgment—addressing one of Tesla’s most critical weaknesses, the lack of basis for judgment. It’s a very simple structure, but putting this simple structural change into Tesla is impossible because the source would have to be rewritten from the beginning. That’s why Nvidia is approaching it a bit differently, and not just Tesla, but trying to apply Vision-based methods to other companies as well.
Current Autonomous Driving Industry Structure
For now, I’ll stop here and look at it more next time. The autonomous driving industry as I’m looking at it right now is, simply put, at this level. If you think of Tesla, China, and Nvidia’s general-purpose architecture as a three-strong structure, that’s it.