Assuming your argument is that all you need is cameras, at what point does it become entirely a software problem? Once you have cameras that cover a sufficient FoV around the car? Or can pivot so that they can, like human eyes? How do you know that the cameras you have meet that requirement?
Even then, I'm not sure how you can say its always just a software problem when no one has accomplished it yet.
Why would you think that? If they have much better technology, that they are deploying on public roads, why wouldn't they let the public know with easily accessible proof?
Obviously this is cherry picked, but you don’t think a car equipped with LIDAR, imaging radar, more cameras, more compute, and temporally dense HD maps that has focused on the same area for its entire existence would perform better than a Tesla?
I would expect a strictly mapped system would perform significantly worse in an area like SF, where roads and obstacles frequently change. Human drivers are not rigid, they are adaptable, so building rigid AV systems is doomed to fail outside perfect conditions.
There are cameras inside of the cars in AVs that face towards the passenger area - it's trivial to have someone check for crud in the back of the car before rides and send it back to the depot for cleaning. And automating that via ML would be easy considering the simple environment.
Why do you think the cost of building the HD maps is prohibitive? How much do you think it would cost to say, keep a square mile of a dense urban HD map updated for a year? It's probably way lower than you think.
I didn’t say it was prohibitive but it’s certainly not cheap. The cost is per km rather than per square km. DeepMap charges $5k / km for example (the big companies don’t publish their costs, but it’s widely known in the industry to be on par with this). I can’t seem to find the stats again, but map invalidation happens on the order of weeks to months.
I work in the space - whoever wrote this clearly isn’t familiar with how Waymo, Cruise, Zoox, etc. actually function. Just going to address a few particularly egregious takes:
>Waymo relies on 3D digital maps.
The cost of HD mapping is vastly overstated in popular press and blogs. It’s not at all a limiting factor. You can update it using the same cars that are part of your service. I've seen hard numbers that back this up.
Google’s unable to keep street view up to date because they have a fleet of hundreds of mapping vehicles spread throughout the entire world, vs hundreds of cars in a small service area that Waymo has.
>Waymo is LIDAR depdendent.
This has been discussed to death - the weather point applies to both cameras and LIDAR, and you can train ML models to deal with noise from weather the same way you can with models running on cameras. Cost has also been addressed a million times before- it’s coming down drastically.
>Waymo is GPS dependent
This one is particularly inane.
HD map based AVs localize themselves to the HD maps with their sensors - they’re not actually GPS dependent. The 2-4m of accuracy GPS has is not enough for AVs.
>Internet connectivity issues
>Waymo depends on 3D HD maps. Therefore, the vehicle needs to have onboard downloaded maps at all times. Waymo must download these maps in realtime, or you have to download these 3D maps on your Waymo vehicle before you visit a new Area.
I don't know why the author thinks you'll be able to drive robotaxis from city to city. Regardles, AV companies design their cars so that they don't need a continuous internet connection to function. They know cell connections can be spotty or blocked while going through a tunnel. See https://readwrite.com/2017/01/11/google-waymo-security-tl4/
>No Viable Business Strategy
>Waymo’s primary business strategy is to sell its technology to the vehicle manufacturer.
This is just wrong and I have no clue where the author got that from. The model is robotaxis, and maybe partnering with trucking fleet operators to outfit their cars.
>Simulated driving is not real world driving.
The data stuff has been discussed before in detail but suffice it to say AVs generally aren’t a data problem - they’re a robotics problem.
>To be successful Waymo has to work very well not only in the USA but also in Canada, Mexico, etc
The revenue from ridesharing services is hyperconcentrated in a few dense urban areas in the US. If you got a robotaxi service working well in just a few of those tiny areas, you could capture a huge amount of revenue to fund tackling other areas.
Thank you for an informed take. I agree completely. There is no fundamental reason why Waymo can't work. Sure, there are challenges, but is servicing cars REALLY something Google can't solve?
Even then, I'm not sure how you can say its always just a software problem when no one has accomplished it yet.