atroyn

Thanks for the recommendation, ordered. Hartley is very valuable to know how we got to where we are, from hand tuned features and RANSAC to more modern full photometric optimization and ML. It answers a lot of "why don't we just... " type questions while still providing the foundations.

In Robotics / Computer Vision:

- Thrun et. al, Probabilistic Robotics

- Multiple View Geometry, Hartley / Zisserman

- An Invitation to 3D Vision, Ma

- Pattern Recognition and Machine Learning, Bishop

-Convex Optimization, Boyd

This is a complex question. As you pointed out, having a high fidelity map helps to identify the part of the environment that change, creating 'deltas'.

Another reason is that autonomous vehicles take data from a wide variety of sensors, including cameras, RADAR, LIDAR and others.

None of these modalities is perfectly, and there is always ambiguity and drift. High fidelity, high resolution maps provide a strong prior that helps to resolve these issues.

Besides the sensing issues, high fidelity maps also provide priors with respect to planning and prediction of the behaviors of other road users. A busy intersection should be approaches differently to a small backroad.

In the end, autonomy needs both real-time processing and high fidelity maps to perform effectively.

Somewhat different to the suggestions so far, but Thrun et. al's 'Probabilistic Robotics' is a very good applied probability text, with a focus on physical systems.

Plenty of worked examples and problem sets as well.

OP here. It's tongue-in cheek, in the same way that Mike Judge's "Silicon Valley" is tongue-in-cheek. Please don't hurt me.

Have you ever worked for any of Musk's companies?

These are all questions to be addressed, along with 'how many of those miles should we upgrade', 'what will we get out of it', and 'how do we make a self-driving car'.

There are many many startups working on some fraction of the autonomy stack (e.g. Aurora) or the entire vertical (e.g. Zoox). I call out several specific ones in my comment if you care to read it through.

I work in autonomous vehicle R&D, and if we had an environment filled with reliable sensors it would considerably simplify our problem.

There is no such thing as a _complete, detailed 3D map of driveable areas_. Most current efforts rely heavily on high resolution, large scale, semantically labeled maps, but on any given stretch of road this is only a first approximation of the environment. Live sensors embedded in infrastructure that could pass along real-time information and updates, particularly from directions our on-board sensors can't capture, would be very useful indeed.

And yes, road markings are part of the semantics we use.

Without revealing too much, German automakers have expertise primarily in logistics, with some expertise in automotive design.

There is a vast pyramid of Tier-1 and Tier-2 suppliers that feed into that logistics chain. A lot of those suppliers are promising automakers that they can deliver either components that feed into autonomy, or else autonomy in full. This fragments the effort across literally hundreds of small teams, many of whom have little to no expertise at all in the necessary disciplines.

Additionally, every German automaker also has its own in-house autonomy team, and the relationships between in-house, contracting, and supplier teams is chaotic. Many managers see autonomy as another way to build a small empire and make a name for themselves, and it just results in a huge organizational snarl.

The fact that this is an area no one knows how to execute on yet since it's brand new, coupled with these gross inefficiencies, will mean autonomy efforts from the automakers will be stillborn. They will eventually learn from it though, acquire the right teams, and get on with putting it into their cars just as they're doing with EV.