I believe Mobileye also uses some human markup of scenes, to help train their systems. That could be a good "moat" against competition. It's time-consuming and labor-intensive to duplicate, which discourages others from trying to duplicate it. They could, but it'll be tempting to try to replace it with more software — and that may not be easy either. I've heard the claim that deep learning already beats humans at some tasks, but that's specific to a benchmark. In real life I suspect that humans using a well-designed quorum-based approach can still be more accurate.
Post-Mobileye, my understanding is that Tesla is using various
NVIDIA products for different purposes:
- FSD demo video(s) used Drive PX2 in the car, with multiple cameras, trained using DriveWorks. This would have been the fastest path to a FSD demo.
- HW2 cars come with Drive PX2 and multiple cameras. But AP only seems to use the front cameras, possibly just one of them. I suspect it's trained using Tesla's own software (Tesla Vision) instead of NVIDIA's DriveWorks, because Tesla wants to own that know-how.
Either or both of these could have been trained on the NVIDIA DGX-1, or on commodity hardware. Either or both could use NVIDIA DIGITS, or other software. Tesla has probably evaluated different options and uses the more cost-effective approach.
Now I understand Tesla has released AP for HW2 at up to 55 mph. Has anyone else demonstrated speeds above 55 mph with PX2 hardware? The Tesla FSD demo videos seemed to top out around 25-35 mph, and the NVIDIA demo at CES seemed pretty slow too. What about Volvo? Aren't they using PX2 in limited trials in and around Gothenburg? How fast can it go?