Waymo brings in $2.25 billion from outside investors, Alphabet

[diplomat33]

Thanks for all the information. This is the future and I am very interested to learn. I am a newbie on this but the common sense tell me one needs two critical components to make the AD happen: Sensor system to get information (hardware) and then the AI (software) to tell car how to drive based on the information. For me, the software will be the key. As for the sensor system, more accurate information it can gather, the better. So I do not understand why Tesla insists on cameras only, unless it adds on too much cost otherwise. The question is, why Waymo can do L4 now but Tesla cannot? Is it because the hardware or the software? If it is just software, then Tesla should be able to catch up with all the data gathered from all those AP drivings. If Waymo can do it, why cannot Tesla?

The software is by far the most difficult part of the problem. What would be the point of shipping expensive hardware when you're not even sure if it will be good enough until you write the software? Waymo is making prototype vehicles while Tesla is making production vehicles. Tesla's approach is extremely unlikely to work but it's not really costing them any extra since you need all the cameras and compute hardware to do driver assist features anyway.
While Waymo is operating a very limited number of L4 vehicles without a safety driver it's certainly not clear yet that they're as good as a human driver and ready to released in more markets.

I agree the software is the hardest part. Perception is actually easy now if you use hd maps, cameras, radars and lidars. Planning and driving policy is the hard part.

But I do think that sensors and hardware are critical because they determine what data your autonomous car gets that it uses to drive. Without the right data, everything else is pointless because your car will essentially be blind or partially blind. With camera, radar and lidar, Waymo cars are collecting enough data to do perception. Of course, they also worked hard on the software and have advanced driving policy. But with Tesla, they are only collecting camera data which is far less data than Waymo. So Tesla needs to completely solve camera vision in order to do perception. So I think the sensors that Waymo uses, are one big reason why Waymo has L4 now and Tesla does not have L4 yet.

 

[Daniel in SD]

True. There's no question about the above facts at all.

Tesla's system still needs human driver to assist it while Waymo can now have no human driver at the wheel at all in Chandler, AZ.

It's a trade-off between Tesla's L2 generalized solution versus Waymo's geofenced L4 system.

I think this is a fundamental difference in viewpoint. There is a huge gulf between L2 and L3-5. One is a car that drives itself and one is a car that assists a human in driving. They're not really comparable.
We spend hundreds of billions of dollars maintaining the roads we have so that people can drive cars on them. Yet it seems inconceivable to people that we'd spend a small fraction of that to map them so that cars can more safely drive themselves?

 

[Tam]

...map them so that cars can more safely drive themselves?

Around here, roads are mutilated by constructions very often. Someone seems to have a great idea to burry utilities under the city roads and work crews keep digging them up throughout the year.

Constructions seem to be perpetual in Los Angeles and Bakersfield Freeways. GM Supercruise L2 would turn itself off and become useless for those who drive through construction zones.

...There is a huge gulf between L2 and L3-5...

Agreed.

However, Audi AI traffic jam pilot L3 works well as long as the speed is 37 MPH or under.

Waymo L4 works great as long as I am willing to sell my house to live in Chandler, AZ.

So, at current, for consumer choice, I've picked Tesla generalized L2 any time!

 

[Daniel in SD]

Around here, roads are mutilated by constructions very often. Someone seems to have a great idea to burry utilities under the city roads and work crews keep digging them up throughout the year.

Constructions seem to be perpetual in Los Angeles and Bakersfield Freeways. GM Supercruise L2 would turn itself off and become useless for those who drive through construction zones.

Cruise's L4 prototype vehicles work in construction zones in their demo videos.

However, Audi AI traffic jam pilot L3 works well as long as the speed is 37 MPH or under.

It was never released (here or in Germany).

 

[mikes_fsd]

Cruise's L4 prototype vehicles work in construction zones in their demo videos.

Tesla's Autonomy Day demo worked flawlessly... they even reported it to the Cali DMV.

Demo's are not going to intentionally show your product failing.

We can agree probably on the following:

• Vision - not just seeing but understanding what you're seeing - (the software) is the most important and the hardest part of the FSD problem.
• You need vision for Camera's and vision for Lidar (if you use Lidar)

Tesla's approach is that Lidar is not needed.
My point is that you're wasting resources by relying on Lidar because you still have to solve the vision problem, but now you are doubling your problem because Lidar vision software does not work on camera vision software.

 

[Tam]

...wasting resources by relying on Lidar because you still have to solve the vision problem...

I think the LIDAR folks think that it's true that vision software needs to be solved but they don't want to wait. Instead, they would use a training wheel, a "crutch" or LIDAR.

LIDAR software is readily available right now so there's no point waiting for the perfection of vision software.

 

[Daniel in SD]

Tesla's Autonomy Day demo worked flawlessly... they even reported it to the Cali DMV.

Demo's are not going to intentionally show your product failing.

We can agree probably on the following:

• Vision - not just seeing but understanding what you're seeing - (the software) is the most important and the hardest part of the FSD problem.
• You need vision for Camera's and vision for Lidar (if you use Lidar)

Tesla's approach is that Lidar is not needed.
My point is that you're wasting resources by relying on Lidar because you still have to solve the vision problem, but now you are doubling your problem because Lidar vision software does not work on camera vision software.

I've always said here that demo videos are pretty useless. Tesla's demo was especially dumb since it just showed driving in trivially difficult situations.
As far as we can tell from the disengagement reports Cruise and Waymo are not having problems with perception. I think it's a mistake to assume that computer vision will be "solved" any time soon.

 

[mikes_fsd]

LIDAR software is readily available right now so there's no point waiting for the perfection of vision software.

Lidar vision software is definitely not readily available. That is why there are so many papers and articles on solving the raindrop/snowflake/fog issue with Lidar.
That is why this crutch is useless because without additional algorithms (what I keep calling vision) Lidar is only really helpful in good weather, where cameras are good at anyway.

I think it's a mistake to assume

Not really making assumptions, but like I said earlier, placing bets.
When I purchase a Tesla with FSD, my bet is on Vision without Lidar. (I would guess that most people who prepay for FSD are making that same bet)
Google/Waymo bet on Lidar!

Computer vision is not new, what is "new" is applying it at such scale where the entire frame (and every frame) is analyzed+labeled for EVERYTHING, not a few items then using the sequenced frames to infer actions of surrounding objects and much more.
And that is all before the car actually acts on that information, this is just the VISION part not the driving part.

So, the bet here is that the hardware and the software that Tesla is working on is enough to start solving the problem of vision in a meaningful way.
And the beauty of the entire Tesla ecosystem is that the entire fleet gets better as the neural network sees more disparate driving scenarios across the fleet and learns how to deal with them.

@diplomat33

 

[Tam]

...When I purchase a Tesla with FSD, my bet is on Vision without Lidar. (I would guess that most people who prepay for FSD are making that same bet)...

I bought FSD for my current 2 cars but not because I think Tesla will accomplish FSD in my lifetime but because it's the only logical choice for consumers currently.

I can't wait for Waymo, I can't use GM Cruise in the city, it's not practical for me to drive on the freeway with a speed of 37MPH for Audi AI traffic jam pilot...

...Computer vision is not new...

Not new but it is still in its infancy or science fiction. That's why some cities ban facial recognition:

UK police's facial recognition system has an 81 percent error rate

Imagine that the car would brake or get into an accident 81% of the time and drivers can have a stress-free drive 19% of the time!

...raindrop/snowflake/fog issue with Lidar...

It's true that LIDAR has not solved the issue of bad weather yet. That could be a limitation that can never be solved.

However, please note that all Autopilot deaths so far happened in good weather, good and clear vision.

First thing first, can a system work reliably in good weather just yet?

 

[wuxyjj]

I just watched a test drive video on "Tesla Autopilot Vs Openpilot"
Self-Driving Test: Tesla Autopilot ($7,000) Vs Openpilot ($1,000)
Seems a cheap aftermarket kit with open source software can do very good job as standard AP function on Tesla. I am sure that Tesla's software development is much more advanced than the open source, but the end results are at the same level. That tells me that depends on software development alone is too hard (or hopeless). What is the point to save $500 for a Lidar system if it can provide extra information to make it easier to write drive polices. Why cannot one combine Cameras, Long Range Radar, Short & Medium Range LiDAR, Ultrasound, GPS, to create real time HD maps? Tesla's car already had everything except Lidar. Waymo's car can self drive in AZ because they know everything around them.

 

[mikes_fsd]

Then Tesla is not right for you (at least not while Elon Musk is in charge).

Companies have combined "Cameras, Long Range Radar, Short & Medium Range LiDAR, Ultrasound, GPS" but NO one - not even Google after 12+ years of Waymo - has figured out how to make the package cost effective and manufacturable at scale.
Waymo is on Gen5 of their hardware suite and with every release, they state "now it is scalable".

We will not see LIDAR standard on any Tesla vehicle. So, if you purchased a Tesla with FSD and expect to see LIDAR somehow get thrown in the mix, you will be sorely disappointed.

Good news! You do not need LIDAR for safe, reliable and economical autonomous driving.
Computer Vision (understanding of your environment at all time) is the solution and that solution is closer then the alternative of LIDAR Vision PLUS Computer Vision.

 

[diplomat33]

Good news! You do not need LIDAR for safe, reliable and economical autonomous driving.
Computer Vision (understanding of your environment at all time) is the solution and that solution is closer then the alternative of LIDAR Vision PLUS Computer Vision.

There is no evidence for this at all. Nobody has safe, reliable and economical autonomous driving with cameras only. Nobody.

In fact, the companies with lidar vision + computer visioon like Waymo are closer to safe, reliable and economical autonomous driving than Tesla is. So the evidence would suggest the opposite actually.

 

[mikes_fsd]

Nobody has real autonomous driving PERIOD
I am sorry that I called your baby ugly, but LIDAR is a "fool's errand"

 

[diplomat33]

Good news! You do not need LIDAR for safe, reliable and economical autonomous driving.
Computer Vision (understanding of your environment at all time) is the solution and that solution is closer then the alternative of LIDAR Vision PLUS Computer Vision.

So when do you believe that Tesla will achieve safe, reliable autonomous driving (total robotaxi, no driver, anywhere in the US, coast to coast)?

 

[Thp3]

Standard future guessing is to expect too much short term and drastically underestimate long term change.

So, I expect in 2020, good weather, good roads, driver in driver seat texting on phone, to be safely driven home from a night out. Maybe I’ll be able to read a kindle book while it drives me home. But weather and roads will be key limits.

in 2023, I expect robotaxi behavior in good weather, maybe 2022. In 2023, with bad weather, the car might tell me “no, it’s not safe to drive at the moment”.

 

[mikes_fsd]

I expect in 2020, good weather, good roads, driver in driver seat texting on phone, to be safely driven home from a night out. Maybe I’ll be able to read a kindle book while it drives me home. But weather and roads will be key limits.

in 2023, I expect robotaxi behavior in good weather, maybe 2022. In 2023, with bad weather, the car might tell me “no, it’s not safe to drive at the moment”.

I like this timeline, but I do think that once the base functionality "in good weather" is working, the timeline speeds up exponentially ...
So, move all the other years left by one year ... good weather Robotaxi by end of 2021.
Bad weather by end of 2022.

But if you really want to know my actual deadline (whether or not I made a good bet on Tesla)...?
June 2025 - my daughter gets her drivers/learners permit.

 

[MP3Mike]

So when do you believe that Tesla will achieve safe, reliable autonomous driving (total robotaxi, no driver, anywhere in the US, coast to coast)?

So when do you believe that Waymo will achieve safe, reliable autonomous driving (total robotaxi, no driver, anywhere in the US, coast to coast)?

My answer is I have no clue for either, but I give Tesla a 60% chance of getting there before Waymo.

 

[diplomat33]

So when do you believe that Waymo will achieve safe, reliable autonomous driving (total robotaxi, no driver, anywhere in the US, coast to coast)?

My answer is I have no clue for either, but I give Tesla a 60% chance of getting there before Waymo.

I expect Waymo to get there in 1-3 years, definitely before Tesla does.

 

[mikes_fsd]

I expect Waymo to get there in 1-3 years, definitely before Tesla does.

What does that even mean in Waymo terms?

With Tesla we know once it is available and you paid for FSD you can verify it independently to your heart's delight. But with Waymo what would them "getting there in 1-3 years" look like?

More locked down routes in more cities?
More cities but not locked down routes?

How does one verify Waymo's claims about its fleet?

 

[MP3Mike]

I expect Waymo to get there in 1-3 years, definitely before Tesla does.

Is the 1-3 years how long you think it will take Waymo to perform the required HD mapping of the entire US? (Which of course will never be done.)