Lex Fridman interviews Chris Urmson

[mblakele]

Lex Fridman's latest AI podcast is an interview with Chris Urmson, formerly of CMU and google/waymo, now with Aurora.

audio: https://media.blubrry.com/takeituneasy/p/content.blubrry.com/takeituneasy/mit_ai_chris_urmson.mp3

video:

There's some Tesla-specific content: Urmson argues against ADAS/L2/L3 as a safe or productive avenue for development. He also argues that, while camera-based driving has an existence proof, lidar is the better path because development goes faster and the increased value of L4/L5 will easily cover the cost of the sensor.

 

[diplomat33]

It seems the camera vs lidar debate basically boils down to this: cameras will take longer to get to L4 but when they do, you'll have a system that works and is cheap or lidar which will get you to L4 faster but is currently more expensive. Both approaches have merit.

 

[EVNow]

There's some Tesla-specific content: Urmson argues against ADAS/L2/L3 as a safe or productive avenue for development. He also argues that, while camera-based driving has an existence proof, lidar is the better path because development goes faster and the increased value of L4/L5 will easily cover the cost of the sensor.

Its difficult to separate a person's view from what his/her company does.

Now, the company might be doing something because the CEO thinks that the absolute right thing to do or that's the best option for that company given the constraints. But the CEO will always say is the correct thing (never say its best for the constraints of the company).

 

[EVNow]

BTW, guess who says they will be "soon" feature complete ? ("few" months from 9/18)

How Self-Driving Supergroup Aurora Plans to Make Robocars Real

In the next few months, Urmson says they should be nearly “feature complete”—capable of doing everything a human driver can, if with less skill. After that, he says, it’s a matter of improving each ability.

 

[diplomat33]

Urmson argues against ADAS/L2/L3 as a safe or productive avenue for development.

I disagree with Urmson on this point. Yes, just adding more and more L2 features on top of each other will never get to you L4 autonomy. But I don't think that is what Tesla is doing. Tesla is developing their vision neural net step by step to eventually cover the entire OEDR. Once you do have the entire OEDR, your car will be a self-driving car. The nags are just a safety safety measure until the car does become FSD. For Tesla, the gradual L2 to FSD approach makes sense because it allows Tesla to monetize AP and deliver value to their customers while they are waiting for FSD. Tesla needs to sell cars and has a fleet of hundreds of thousands of cars already on the road with L2. Plus, Tesla can roll out features and get feedback from the entire fleet to help perfect the feature.

But for companies like Aurora that don't have to worry about selling cars, they can start with a self-driving prototype and focus on FSD without worrying about L2.

Having said that, I do get his point about driver overconfidence.

BTW, guess who says they will be "soon" feature complete ? ("few" months from 9/18)

How Self-Driving Supergroup Aurora Plans to Make Robocars Real

That sounds very familiar. I wonder where have I heard that before.