- First principles hardware design of focused self driving computer (many times better than any competing existing hardware). Already shipping in all newly produced cars. Currently working on next gen that will be 3x better to ship in a couple years.
- Lidar is an unnecessary mistake that competitors are making that won't succeed (too expensive, need too many, unnecessary).
- Real world fleet testing is critical to success, simulations are not good enough since there are too many unknown unknowns in the real world. Tesla uses simulations too, but nobody else comes close on real world fleet testing.
> Lidar is an unnecessary mistake that competitors are making
This is very controversial and rest of the industry thinks exact opposite, many even claiming that Tesla is being irresponsible and even delusional in trying to do autonomy without lidar. The main points I have heard in favor of lidar are that computer vision is very flaky not only in suboptimal weather but even in good weather. Cameras are simply no where close to in performance in dynamic range, rapid adaptation, focusing etc as human eye. Imagine car going under the shaded road with rapid changes between bright light and shade. The likelihood that your depth estimation will get messed up is very high. Of course, night driving becomes highly questionable as well. In additional the long range depth estimation is very flaky with stereo vision right now and a topic of research for mono-vision. If you want to retreat to level-4 only and that too with conservative speed, weather etc then may be vision+radar more doable?
They pair it with a forward facing radar which is inexpensive and good at depth perception.
Elon predicts all competitors will eventually drop lidar. He mentions it's expensive, but also not as good in a lot of cases (and all roads/signs are designed for vision).
He argues that getting vision to work is a prerequisite for getting self-driving to work and once you have it working, lidar is worthless (and unnecessary).
Elon Musk is noted for his premature optimization.
Case in point: both the Gigafactory (vastly overbuilt for the quantity of batteries actually produced) and the Alien Dreadnought (vastly overbuilt for the number of cars Tesla current produces...assuming that Tesla is ever able to get the fancy automation working). Boring Co digging a two-mile tunnel in West LA without bothering to learn how to pour concrete smoothly, or to make the "rails" the proper width, or learning about ventilation, or access points....
Often reverse is true. Trying to optimize system with million lines of code may require serious change in architecture and much harder because of backward compatibility and all the legacy baggage. As many people would describe it, it much harder to fix an airplane that must also continue to stay in the air.
It's not easier to optimize a working system for consumers if "optimization" == "not dying". What will happen is you will deoptimize your brand's safety and find yourself regulated.
“In my view, it’s a crutch that will drive companies to a local maximum that they will find very hard to get out of,” Musk said. He added, “Perhaps I am wrong, and I will look like a fool. But I am quite certain that I am not.”
Their forward facing radar has failed to avoid crashes into big, stationary objects (at least 2 semi trailers, a fire truck, and more). The explanation I got was that radar is rather noisy, so it's handy to filter-out stationary objects (like road signs, broken down vehicles on the shoulder, and fatally, any any semi trailer crossing your lane)
I think it also fundamentally depends on the business model being pursued:
- If you are trying to sell the Model 3 directly to the end consumer with autonomous mode, the extra $10K for Lidar and bulk (which will greatly effect the exterior design) are definitely non-starters.
- If you are going to robo-taxi route (such as Waymo and Uber), the extra one time $10K cost to add lidar for the 5 year life of the car is probably a blip on the income statement of the operator as compared to a full time human driver which probably costs $10K PER MONTH for the lifetime of the service. For the robo-taxi business model - its a bit of a no brainer - they could stick every sensor known to man on the car and still make out like a warlord by getting rid of the human driver but maintained a 100% safety record. Plus making the car stand-out with a unique Lidar inclusive shape is a great marketing differentiator. Also reduces your liability if a taxi rider ever sues since you can claim you have redundancy in the system.
I don't think anyone wants to use Lidar - the claim is that they need it to achieve full, safe, autonomy. If it turns out that it isn't needed than Tesla will be in a great position.
Not that I'm confident about either side of this, but the TOPS isn't the only important factor here. It's also the efficiency. If Tesla's CPU can do more ops/watt, then it's more ideal for the conditions that Tesla needs the CPU to fulfill. On top of that, I have no idea what "320 TOPS" means in the context of the specific workload Tesla has. Will they _actually_ get 320 TOPS?
Eh, I'm pretty fine with that. Imagine the current gen slightly beats human drivers - also known as "revolutionary, but not perfect" - of course the next gen would focus on improving safety further.
> Currently working on next gen that will be 3x better to ship in a couple years.
That's ... not great. If that was CPU type of unit, it'd be great. But TPU-type accelerators are growing at massive speed (as it's still pretty new and simple tech), where you're more looking for 10x type of gains.
- First principles hardware design of focused self driving computer (many times better than any competing existing hardware). Already shipping in all newly produced cars. Currently working on next gen that will be 3x better to ship in a couple years.
- Lidar is an unnecessary mistake that competitors are making that won't succeed (too expensive, need too many, unnecessary).
- Real world fleet testing is critical to success, simulations are not good enough since there are too many unknown unknowns in the real world. Tesla uses simulations too, but nobody else comes close on real world fleet testing.