My Authors
Read all threads
How can Tesla progress from its current Autopilot product to Robotaxis? (continued - part 2)

1) What about the staff bandwidth bottleneck?
Tesla’s access to driving data is critical in my view, but it isn’t a magic bullet.
2) It still takes a lot of human input to solve each problem & Tesla has to first identify the most pressing problems/ important driving scenarios to solve & slowly work its way through the list towards rarer & rarer problems.
3) This is the march of 9s. The continuous quest to get probability of zero disengagements per mile to the next decimal place.
4) Tesla’s fleet drives ~1 billion miles per month, but most cars are only rarely collecting data because most driving situations encountered are common, easy & already solved.
A key part of the challenge is working out when to collect data.
5) I believe this is likely decided by collecting high level data on Autopilot disengagements & “shadow mode” disagreements between Autopilot’s plans & the human’s actual actions.
6) This can be used to form a priority list of the most common remaining problems that will make the largest difference to fleet disengagements. Once these problem scenarios have been identified, the car fleet will be told to look for these situations & collect full sensor data.
7) Once data is collected the image frames have to be labelled by human annotators. This data can then be used to train neural nets to solve the driving scenario & to create test sets to test new iterations of the Autopilot software.
8) This is why working through the list of driving scenarios/problem cases takes time & Tesla’s theoretical access to data doesn’t instantly solve self-driving – they still have a finite level of human resources to work on each problem.
9) Tesla’s solution to speed up progress on the march of 9s & to solve the staff bandwidth bottleneck: “Operation vacation” & 4D partially self-labelled neural nets.
10) Tesla’s team has been building infrastructure to automate the identification of priority driving scenarios & to automate fleet requests to collect data to solve these issues.
11) They are also likely automating parameter optimisation of neural nets & automating the process for testing these against a set of test scenarios to choose the best performer.
12) Additionally, the move to 4D video based neural nets should also dramatically increase labelling efficiency as most connected frames/objects across the 8 cameras can now be labelled automatically after initial human input.
13) So, Tesla’s Autopilot capabilities have been held back by two key bottlenecks. Both of which Tesla is potentially on the verge of solving (but as i said in my first thread, for now i still put this at low probability).
14) 4D neural nets may solve pseudo lidar & a largely automated solution towards solving the March of 9s can lead to faster & faster reduction in disengagements.
15) If this works then before long disengagements will be rare enough that Tesla will be able to collect full sensor data from the fleet for every single disengagement -simplifying the data collection filter problem.
16) If this all goes to plan will there be further bottlenecks along Tesla’s path to 2-10x human driving accuracy? They need more compute, they may need fundamental algorithmic breakthroughs, but just how close they are to a solution should become a lot clearer before year end.
17) Finally, Its worth noting quickly that even with less intelligence & less understanding of causation & motives of other road users, a Robotaxi has many advantages vs humans that could still lead to superior safety – instant reaction times, constant 360 degree attention etc.
Missing some Tweet in this thread? You can try to force a refresh.

Keep Current with Reflex Research

Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Follow Us on Twitter!

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3.00/month or $30.00/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!