Umm Waymo need help presumably - design the road around the car... Scalability is off the charts...
An Alphabet company is designing a road for autonomous cars in Michigan
So this is why you don't think FSD is a solved problem
@diplomat33 ?
If you are implying that Waymo is helping with this project because they cannot solve FSD, no I don't think so. This project is separate from Waymo's own efforts to solve FSD. This is a project by the State of Michigan. There are those who believe that our poor and inconsistent road infrastructure is making autonomous driving harder than it needs to be. So they argue that perhaps we should try to design roads that are more friendly to autonomous cars. That is what this project is all about. Waymo is helping with the project because they are a leader in FSD so they certainly have expertise to contribute in terms of how to make roads more friendly to autonomous cars. There is no doubt that if we had special roads designed just for autonomous driving, that we could make autonomous driving a whole lot easier, that's for sure.
But in terms of why FSD is not "solved" yet, the perception part of autonomous is solved with camera+radar+lidar+HD maps. And Waymo has solved the perception part. Waymo also has advanced planning/driving policy, hence why they are able to deploy robotaxis with no drivers in geofenced areas. The part that is not solved yet is the prediction part of autonomous driving. Predicting the behavior of other vehicles, pedestrians, especially when there are a lot of them, can be a challenge. Consider that some cars forget to signal, some pedestrians change their minds at the last moment, cyclists may forget to signal etc... So you need to predict possible behavior of agents that don't always signal their intent and sometimes behave in an unpredictable way. So "solving" the prediction part may not be possible. But the goal is to train the autonomous car to at least anticipate the possible behaviors of all agents as accurately as possible so that the car can drive as safely as possible.
I base this on what Waymo's CTO said. Last year, in an interview with TechCrunch, he said that:
"Detecting different objects, understanding what they are, classifying pedestrians, cyclists, cars etc... It gets super interesting beyond that. That's step 1. Self-driving 101, you have to see the world and understand the different types of objects. Really where it gets interesting and where we have been doing a lot of fundamental work on the research front and the engineering, is reasoning about the world at the scene level, not just the per object level, and really deeply understanding the interactions between the different objects in the world. How does one car fit with the pedestrian or the cyclist? How do you fit in to that? This very deep, very dynamic, very social interaction, is where a lot of really interesting work happens."
Source:
Lastly, I would add that "solving" L5 is exponentially harder than "solving" L4. That's because L5 involves an absurd amount of edge cases and driving cases to deal with. Getting the autonomous car to handle ALL the cases in L5 reliably is a monumental challenge. "Solving" L5 may not be possible, at least not right away. Waymo has FSD that works but if they deployed it everywhere, it probably would not be reliable enough to remove the driver because of how many different cases it would need to be able handle. That's why Waymo focuses on L4. With L4, you can limit the cases you need to deal with. Less cases makes it easier for the FSD and it makes it easier to validate when your FSD is ready to be deployed without a safety driver. Hence, why Waymo has been able to remove the safety driver for some geofenced routes.