Autonomous Car Progress

[EVNow]

Some interesting things the Cruise CEO said just a few months back that is opposite of what Waymo fans here insist is the "only way" to make FSD work. And he acknowledges that using multiple sensors, HD Maps and Geofencing are all temporary measures and the ultimate goal is not achievable with these.

04:26 : okay so let's take a step back and just expand a bit more on our overall approach to autonomy the goal obviously in the medium to longer term for everybody working on this is to get to low-cost generalized autonomy we want fully autonomous driving we want it to work everywhere and we want it to be really inexpensive

04:48 : so there are a couple different ways that you can approach that problem you can start with a really low cost system like an L2 system and try to work your way up from a performance point of view or you can take the approach that we've taken

05:31 in order to have a shot at solving that we then relaxed a bunch of other constraints so we said we're going to throw as much compute as we possibly can at the problem we're going to throw as many sensors as we need to at the problem we're going to use high definition maps to help get there and we're going to put in place a geofence if we need to sort of optimize around that ...

Wow ... so what do we have here ? No - he is not "gloating" about multiple sensors and HD Maps ... he says they are "relaxing constraints" (a.k.a compromises / crutches they are using). He clearly understands and says they took all these short term measures in order to get to 1000x better at driving faster.
- Lots of sensors
- HD Maps
- Geofencing

But some want the world to believe that crutches are actually great and indeed the only way to get to FSD - looks like the classic "make a virtue of necessity".

 

[diplomat33]

As long as they can show a few million miles of real-world testing ... should be fine. Ofcourse they will learn the hard way that simulation is only as good as the programmer's imagination. And as we all know, programmers never make mistakes

Personally, I think you need real world testing during development. But I am intrigued by the idea of trying to "solve FSD" with simulation only. Simulation is becoming very realistic both physically and in agent behavior. And we've seen AI like AlphaGo learn how to play. So the idea of a driving AI learning how to drive on its own entirely in a simulation, is thought provoking to me.

 

[EVNow]

Personally, I think you need real world testing during development. But I am intrigued by the idea of trying to "solve FSD" with simulation only. Simulation is becoming very realistic both physically and in agent behavior. And we've seen AI like AlphaGo learn how to play. So the idea of a driving AI learning how to drive on its own entirely in a simulation, is thought provoking to me.

Comparing to alpha-go would only be correct if the company tries end-to-end NN. AFAIK, nobody tries that now.

 

[diplomat33]

Comparing to alpha-go would only be correct if the company tries end-to-end NN. AFAIK, nobody tries that now.

Actually, Wayve is doing end-to-end NN. AFAIK, they are the only ones.

 

[EVNow]

Actually, Wayve is doing end-to-end NN. AFAIK, they are the only ones.

Looks like a small startup ? Good luck to them.

But the one claiming to simulate their way to FSD are not end-to-end FSD, right ? BTW, Hotz company is also doing end-to-end … but not trying for full autonomy.

 

[Daniel in SD]

Some interesting things the Cruise CEO said just a few months back that is opposite of what Waymo fans here insist is the "only way" to make FSD work. And he acknowledges that using multiple sensors, HD Maps and Geofencing are all temporary measures and the ultimate goal is not achievable with these.

04:26 : okay so let's take a step back and just expand a bit more on our overall approach to autonomy the goal obviously in the medium to longer term for everybody working on this is to get to low-cost generalized autonomy we want fully autonomous driving we want it to work everywhere and we want it to be really inexpensive

04:48 : so there are a couple different ways that you can approach that problem you can start with a really low cost system like an L2 system and try to work your way up from a performance point of view or you can take the approach that we've taken

05:31 in order to have a shot at solving that we then relaxed a bunch of other constraints so we said we're going to throw as much compute as we possibly can at the problem we're going to throw as many sensors as we need to at the problem we're going to use high definition maps to help get there and we're going to put in place a geofence if we need to sort of optimize around that ...


View attachment 771817

Wow ... so what do we have here ? No - he is not "gloating" about multiple sensors and HD Maps ... he says they are "relaxing constraints" (a.k.a compromises / crutches they are using). He clearly understands and says they took all these short term measures in order to get to 1000x better at driving faster.
- Lots of sensors
- HD Maps
- Geofencing

But some want the world to believe that crutches are actually great and indeed the only way to get to FSD - looks like the classic "make a virtue of necessity".

View attachment 771824

I’ve always said that starting with something driverless that works and then doing successive cost downs is the way to go. Sort of like the Roadster -> Model S -> Model 3 or any other technology.
I think the fundamental difference between your philosophy and Waymo’s is that they don’t think that door to door driver assist has any value at all (and I agree with them).

 

[AndreP]

In case it hasn't been posted in this thread, here's what seems like a fantastic interview with James Douma regarding FSD -- this guy works in AI etc

I feel like Dave came in with a pretty positive bias and James brings it down to reality. James can't put a date on "Robotaxis", the question about releasing FSD to the whole fleet this year is also uncertain and something that he and his technically-minded friends bet about.

The narrative here is now slowly changing to the wide release of "supervised FSD", whether Tesla has incentive to do that, and when it could happen. James also talks about the necessity of flipping to a Waymo-style heavily geofenced setup for the next steps towards what people picture as full autonomy and Robotaxis.

 

[voyager]

The intervention from the NHTSA is one of the first we’ve seen from the agency, which has taken a fairly hands-off approach to regulating the development and testing of self-driving veh dating back to the Obama administration.
“Innovation must not come at the risk of public safety,” Heidi King, deputy administrator of the NHTSA, said in a statement. “Using a non-compliant test vehicle to transport children is irresponsible, inappropriate, and in direct violation of the terms of Transdev’s approved test project.”

Update October 23rd, 2:30PM ET: Added information from Transdev’s official statement.

I see Heidi King mentioned as a NHTSA official involved with ADS projects...
Are there more NHTSA officials known here?

 

[diplomat33]

Baidu has launched their robotaxi service in Shenzhen, China. There are 50 different pick up and drop off points for now with plans to grow to more than 300 pick up and drop off points by end of this year.

But this line is curious to me:

For now, customers are still required to sit in the driver’s seat as part of a safety precaution. Baidu promises that this will change in the future as it continues to perfect the autonomous technology.

I assume that is a mistake and the article meant to say "safety driver", not "customer". I can't think of any robotaxis that make the paying customer supervise the robotaxi. That seems odd.

Source: Will you trust a robot taxi? Baidu’s self-driving taxis take on Shenzhen, one of China’s busiest city | WapCar

 

[voyager]

Waabi is a new start-up that thinks that the current approach of developing FSD has proven to be too costly and too slow. They think they can do better with simulation alone, with no real world testing until the very end:

This super-realistic virtual world is a driving school for AI

Startup Waabi is betting everything on a radical new approach to autonomous vehicles that ditches real cars.

www.technologyreview.com

I have read the MIT article. Interesting. Waabi is mentioned.
Are there more companies solely focused om simulations?

 

[diplomat33]

I have read the MIT article. Interesting. Waabi is mentioned.
Are there more companies solely focused om simulations?

AFAIK, Waabi is the only company focusing only on simulations to "solve FSD". Of course, almost all AV companies use simulations but they use them in conjunction with real world testing,

 

[Dewg]

Forgive this ignorant question, but when you're riding in a Waymo and the car needs help from a human driver, does the passenger know that a human is helping at any time (some notification or indicator)?

 

[diplomat33]

Forgive this ignorant question, but when you're riding in a Waymo and the car needs help from a human driver, does the passenger know that a human is helping at any time (some notification or indicator)?

Yes. If you are talking about the safety driver taking over, there is a notification when the autonomous driving disengages. Also, when there is no safety driver, there is a notification on the screen when remote assistance helps the car. We can see an example in this video from JJ RIcks:

 

[Dewg]

Yes. If you are talking about the safety driver taking over, there is a notification when the autonomous driving disengages. Also, when there is no safety driver, there is a notification on the screen when remote assistance helps the car. We can see an example in this video from JJ RIcks:

It would be interesting if Tesla ends up doing something similar with their robo-taxi concept. Have a fleet of humans that can help the car out when there is something it doesn't understand.

 

[EVNow]

I can't think of any robotaxis that make the paying customer supervise the robotaxi. That seems odd.

You mean like rental companies do ?

BTW, that is one of the scenarios I've talked about in the past for Tesla. They just need to be able to get to the customer pickup point without a driver. Then it becomes an ADAS / L3 rental. Somewhat reduces risk.

 

[EVNow]

I’ve always said that starting with something driverless that works and then doing successive cost downs is the way to go. Sort of like the Roadster -> Model S -> Model 3 or any other technology.
I think the fundamental difference between your philosophy and Waymo’s is that they don’t think that door to door driver assist has any value at all (and I agree with them).

They also don't think there is any value in expanding beyond a 2 tiny areas in 4 years. I don't agree with them. Afterall Roadster could drive everywhere - not just in a tiny area of Chandler.

BTW, Cruise is talking about cost & performance. You can also talk about geo & performance - which is the point apparently waymo fans here don't get.

 

[Doggydogworld]

Some interesting things the Cruise CEO said just a few months back that is opposite of what Waymo fans here insist is the "only way" to make FSD work. And he acknowledges that using multiple sensors, HD Maps and Geofencing are all temporary measures and the ultimate goal is not achievable with these.

LOL. He actually said they relaxed the cost and ubiquity constraints to focus on solving the problem. He knows sensor and compute cost will drop by orders of magnitude as tech advances and volumes grow. And he knows HD Maps will become ubiquitous as happened with every other kind of map.

He's saying those who don't use the proper tools are trying to solve the wrong problem. Of course he's just talking his book, and missing the big picture. The real trick is not "solving the problem", but getting customers to write checks. Elon has pocketed 1+ billion in cash from happy Faux Self Driving customers and half a trillion in market cap without solving any problem. That's how you win in today's world!

 

[AndreP]

LOL. He actually said they relaxed the cost and ubiquity constraints to focus on solving the problem. He knows sensor and compute cost will drop by orders of magnitude as tech advances and volumes grow. And he knows HD Maps will become ubiquitous as happened with every other kind of map.

He's saying those who don't use the proper tools are trying to solve the wrong problem. Of course he's just talking his book, and missing the big picture. The real trick is not "solving the problem", but getting customers to write checks. Elon has pocketed 1+ billion in cash from happy Faux Self Driving customers and half a trillion in market cap without solving any problem. That's how you win in today's world!

That's one way of winning over the short term, but cash also flowed like water in 2020-2021.

GM has been around for 114 years, these companies are focused on the long game. The more competition and different approaches taken, the sooner we'll get a ton of value from someone.

 

[EVNow]

LOL. He actually said they relaxed the cost and ubiquity constraints to focus on solving the problem. He knows sensor and compute cost will drop by orders of magnitude as tech advances and volumes grow. And he knows HD Maps will become ubiquitous as happened with every other kind of map.

I don't think that is the right interpretation ... there is no other way to explain why he combined
- Multiple sensors
- HD Maps .... with
- Geofencing

Afterall remember .... 4 months back, GM is already working on and well into their UltraCruise effort that doesn't use HD Maps and the leadership says because HD Maps is not practical for 2 Million + miles.

It is now the industry-wide consensus that HD Maps don't scale beyond big cities. GM says it, ME says it, Tesla says it ...

ps : I don't know whether ADAS can increase performance 100x/1000x everywhere in the next couple of years to become truly autonomous like Tesla, ME, GM, XPeng, NIO are trying to do. I also don't know whether geofenced robotaxi companies can geographically scale quickly in the next couple of years to become truly ubiquitous. So, I remain a sceptic in the short/medium term of both approaches.

 

[ThomasD]

So if HD maps aren't scalable what mapping are they going to use? You can't always rely on GPS