1/ Uber's AV crash was reviewed by the NTSB and they have released their conclusions.

It looks like they were investigating an airline and not an AV company, lets look at the difference.

https://www.ntsb.gov/news/press-releases/Pages/NR20191119c.aspx
2/ The collision scenario in brief:

* Road was straight and flat, weather was clear
* Only the AV and pedestrian were on the road
* Pedestrian was detected on the ranging sensors, hardware worked.

This was a software failure in the face of an extremely simple scenario.
3/ NTSB focused on people process:

* Vehicle operator distraction
* Uber's safety culture
* Non-regulation at the federal and state level

Focusing on people in important for airlines as they have to coordinate 40-50 employees per plane.
4/ Pilots are the most critical employee. They spend many years in training and still have stacks of manuals that remind the pilot how to fly in exceptional cases.

Planes, software, manuals and pilots are all part of the same system but only the pilots are blamed for failure.
5/ AVs operate in a world without manuals or pilots, just ordinary people inside are outside the car.

Any machine has irreducible failures, leaving enormous pressure on the software to be correct.
6/ Cars allow two significant simplifications over planes that make full automation possible:

1. Stopping a car is much simpler than landing a plane.

2. We can observe how people around cars act today.

This implies a significant change in product development.
7/ Observation forms the core of development for AV because people aren't going to change their behavior.

The risk of testing in public requires focus on a particular question.

Safety driver unattentiveness is a symptom of an unfocused driving test plan.
8/ Everything you observe or even choose to look for during testing must end up in software.

Documents and reports don't count.

Testimonies and intents don't count.

Safety officers and lobbyists don't count.

Only what can be repeatedly tested in software, by software counts.
9/ Over 18 months, it's reasonable for the NTSB review 100% of Uber's hardware and software.

Almost all problems in AV will be in software.

Almost all software problems are identifiable from the software testing environment.
10/ Instead of focusing on people procedure, ask:

* What exactly were we hoping to learn from a live test?
* Was there a way to learn without a live test?
* How will this be encoded in software?
* Where do you verify your assumptions in software?
11/ AVs mean the end of pilot error and a change to the core assumptions of product development.

NTSB, other investigation, and regulatory bodies will have to develop high competence in software systems to understand safety in the future.
You can follow @ghosthayes.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: