Tesla's FSD has different approach / tradeoffs compared to dedicated robotaxi services. FSD has to be cheap and energy efficient, run completely on-board, and it must work everywhere. They're trying to do more with less, which has so far been impossible. Their cybercab and robotaxi service will probably work more like Waymo, with a slightly relaxed set of limitations.
Some differences compared to Waymo:
- Waymo has / can use more on-board compute, from [0] "It has also been revealed that Waymo is using around four NVIDIA H100 GPUSs at a unit price of 10,000 dollars per vehicle to cover the necessary computing requirements."
- Waymo uses remote operators. This includes humans but can also have remote compute.
- Waymo's neural network model can be trained / overfit on specific route or area. FSD uses the same model everywhere.
- Waymo's on-board hardware can use more energy, because it's possible to charge the battery between trips.
- Robotaxi services charge customers per mile, so it makes sense to run longer routes which are also easier to drive, i.e. the routing algorithm can be tuned to avoid challenging routes. This would be possible to implement on FSD too, but it seems that FSD drives fastest route.
You'd think the biggest win would be in the middle:
We have an interstate highway system that's fairly well-maintained and understood, and is a finite space to map. Hypertrain on that, and you can offer an experience of 10 minutes hands-on-wheel at the start and end of the journey, and 3 hours of doomscrolling in the driver's seat. The highway miles are the most boring, both from a surprise-hazard standpoint and from a driver's-attention standpoint (there's nothing cool or interesting to see except the trunk lid of the car in front of you)
It offers a nationwide level of service that Waymo's city-by-city rollout lacks, and the chance for route-specific hueristics that Tesla's cameras-and-local-compute might miss.
Waymo specifically claims they never do remote human piloting. The car will present a remote human operator a choice of routes to get out of a situation, and the human will pick one. Remote piloting is way too risky.
Yes, definitely. "remote operator" is a human or an LLM which is able to make high-level decisions (i.e. what to do in a novel weird situation), but doesn't directly pilot the vehicle. Generally speaking, on-board compute is fast and stupid, and remote compute or a human is slow and smart.
I don't think that cars will have SOTA level LLMs running locally for a long time, and it seems that they actually need that kind of intelligence for full autonomy. However, it might also be totally fine if the passenger makes the difficult high-level decisions through a voice interface.
All decisions are made by the Waymo vehicle itself.
The vehicle can ask human remote operators for recommendations or clarification, but the vehicle itself decides whether to use them. Most of the time, the vehicle doesn't end up needing it.
The system though provides a way to let humans create training data for edge cases.
Some differences compared to Waymo:
- Waymo has / can use more on-board compute, from [0] "It has also been revealed that Waymo is using around four NVIDIA H100 GPUSs at a unit price of 10,000 dollars per vehicle to cover the necessary computing requirements."
- Waymo uses remote operators. This includes humans but can also have remote compute.
- Waymo's neural network model can be trained / overfit on specific route or area. FSD uses the same model everywhere.
- Waymo's on-board hardware can use more energy, because it's possible to charge the battery between trips.
- Robotaxi services charge customers per mile, so it makes sense to run longer routes which are also easier to drive, i.e. the routing algorithm can be tuned to avoid challenging routes. This would be possible to implement on FSD too, but it seems that FSD drives fastest route.
[0] https://thelastdriverlicenseholder.com/2024/10/27/waymos-5-6...