DTNS 3532a – Slowing Down Self-Driving

Are the timelines for autonomous car rollouts too optimistic? What are the real-world impediments to their implementation and what would be a more realistic time-frame for their deployment?

Starring Tom Merritt, Sarah Lane, Roger Chang and Rob Reid.


Using a Screen Reader? Click here

Multiple versions (ogg, video etc.) from Archive.org.


Subscribe through Apple Podcasts.

Follow us on Soundcloud.

A special thanks to all our supporters–without you, none of this would be possible.

If you are willing to support the show or give as little as 5 cents a day on Patreon. Thank you!

Big thanks to Dan Lueders for the headlines music and Martin Bell for the opening theme!

Big thanks to Mustafa A. from thepolarcat.com for the logo!

Thanks to Anthony Lemos of Ritual Misery for the expanded show notes!

Thanks to our mods, Kylde, Jack_Shid, KAPT_Kipper, and scottierowland on the subreddit

Show Notes
To read the show notes in a separate page click here!


2 thoughts on “DTNS 3532a – Slowing Down Self-Driving

  1. About cars with autonomous driving features:

    I concur with what Sarah said. If you have a car with self-driving features that allow you to pay less attention to driving, then – yeah – people are going to pay less attention to driving. You cannot possibly expect anything else than that.

    About what Rob Reid said:
    Judgement calls:
    As discussed in the show: human drivers make judgement calls all the time; if you find a one-way street blocked (say, due to construction), and for some reason the entry to that one-way street had not been blocked off, so now you’re stuck, then you can make a u-turn and drive back.

    It’s my understanding that in the case of autonomous driving the decision making process is entirely implemented in the form of neural network machine learning. This form of machine learning learns by example. Humans can learn from just a few examples; humans can generalize from that. Machine learning requires a vast number of examples, covering all of the space of possible scenarios. With a sufficently large and diverse supply of examples machine learning can achieve human level performence.

    In the case of the Tesla cars it is my understanding that the auto-pilot feature is computing away _both_ in cars with auto-pilot enabled and with auto-pilot disabled. (Currently, _every_ Tesla car coming out of the factory, S, X, 3, has the sensor array and the auto-pilot computer onboard). Every time the human driver does something the neural network had computed too the neural network takes that as a reinforcement. Every time the human driver does something different from what the neural network planned to do the deviation is used as feedback. Changes in the weights of the nodes of the neural network are uploaded to the Tesla servers, and changes from all the cars are assembled/mixed/averaged.

    With this kind of machine learning the autonomous car will make the same kind of judgement calls that humans make. It’s not algorithmic, it’s not programmed.

    The problem, of course, is that the autonomous car will be poorly equipped to handle cases that are very rare. The rarer the situation type, the longer it takes to accumulate the number and variety of examples that is needed to for machine learning.

    About the March 23 accident:
    As noted in the show: Tesla has (against NTSB protocol) released information: the car was in auto-pilot mode, and the car had been issuing warnings to the driver to put the hands on the wheel again.

    I assume that the warning ‘Put hands on the wheel’ is in effect a warning that says: ‘find out _now_ whether you need to take over the driving’

    My understanding is that Tesla auto-pilot will alert the driver when the confidence level that it is driving correctly drops beneath some pre-set limit. The Tesla update did not mention _why_ the car was prompting the driver, but that is a crucial bit of information. For instance, if there was diminished confidence level, the why did the car not preemptively slow down?

  2. About autonomous cars (pods)
    I was listening to DTNS 3532A about Tesla, Uber and the concern about autonomous cars , ac.
    1: When there are only ac’s there will be no issues with today’s minor violations. And no one way streets are needed as a ac can communicate with other ac etc.
    2: The best thing with ac today; they have , a) data storage which records of accidents, if any. b) They have cameras, anyone who will ‘play’ a joke with a ac will be on camera. And the accidents too.
    3) The problem with the evolution is not ac , it’s the current drivers.
    – If Uber got a temporary stop, they were alone. Waymo, Nvidia, Volvo etc are moving forward with no reduced speed. Can’t wait.
    Stefan , Virginia.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: