Make your robotaxi predictions for the 8/8 reveal

So Elon says that Tesla will reveal a dedicated robotaxi vehicle on 8/8. What do you think we will see? Will it look like this concept art or something else?

I will say that while this concept drawing looks super cool, I am a bit skeptical if it is practical as a robotaxi. It looks to only have 2 seats which would be fine for 1-2 people who need a ride but would not work for more than 2 people. I feel like that would limit the robotaxis value for a lot of people. Also, it would likely need a steering wheel and pedals for regulatory reasons even if Tesla did achieve eyes-off capability.

So I think this is concept art for a hypothetical 2 seater, cheap Tesla, not a robotaxi.

Could the robotaxi look more like this concept art but smaller? It could look a bit more like say the Zoox vehicle or the Cruise Origin, more futuristic box like shape IMO and seat 5-6 people.

Or maybe the robotaxi will look more like the "model 2" concept:

Other questions:
- Will the robotaxis be available to own by individuals as a personal car or will it strictly be owned by Tesla and only used in a ride-hailing network?
- What will cost be?
- Will it have upgraded hardware? Radar? Lidar? additional compute?
- Will Elon reveal any details on how the ride-hailing network will work?

Thoughts? Let the fun speculation begin!

 

[stopcrazypp]

Waymo uses HD mapping but it's not an inherent dependency. It gives confidence. The better the sensor-based model becomes the less benefit there is to the HD mapping.

HD mapping is a cost burden, so naturally you'd want to get away from that and navigate likes humans with simple maps, sensors and local knowledge.

Theoretically that is the case, but in practical reality when your cars exclusively drive in HD mapped areas as a minimum requirement for operation (which is the case for Waymo, as well as a few different L2 systems) you tend to place higher confidence in it than a system that does not use HD maps at all or has it as optional.

Given Waymo cars are still driving straight into construction sites (didn't realize they still do that, I thought only Cruise does that) apparently all that better sensor-based model still isn't good enough when the road doesn't match the HD map.

 

[Joey D]

That's not what I am trying to say. I am trying to say that MACHINE LEARNING (current techniques) alone isn't likely able to unless there are major research progress.

There are leaps and bounds of progress with machine learning (both software and hardware). You may have heard something about it being the hottest technology field. Deployed compute alone is increasing at an order of magnitude every 9 months to a year.

At least it was in marketing? 2019: Robotaxis next year (except to DMV). Now: Robotaxis never? FSD (Supervised) ie a Level 2 with a large ODD.

Who said "no robotxis never?" It will remain a level 2 ADAS system...right up until it isn't.

Waymo's goal is to develop an autonomy system and deploy it as robotaxis, make money.

That is far less ambitious than Tesla's goal then...which explains the difference in approach.

 

[spacecoin]

There are leaps and bounds of progress with machine learning (both software and hardware). You may have heard something about it being the hottest technology field. Deployed compute alone is increasing at an order of magnitude every 9 months to a year.

Doesn't matter how much compute you have if you don't have the data or the techniques to explain or validate the system. Again, data curation and ensuring regressions is the hard part for safety critical systems in ML. And, in the robotaixs that are available today (Waymo, Cruise, Zoox, Baidu, PonyAI) have spent the last 8-10 years on getting to driverless by carefully adding nines by all means possible.

Again, ponder on why not radiology is unsupervised yet. Still images. Not time critical. A simple CV-application compared to driving, right?

Who said "no robotxis never?" It will remain a level 2 ADAS system...right up until it isn't.

Sure. Good luck with that. I'll go out on a limb and say that no one will deploy driverless vehicles with CV/ML only this decade, and most likely not in the next either unles there are some seriois research breakthroughs.

For sensing, Lower hardware costs beats the sh!t out of ML scaling attempts where you need exponential training data sets for linear progress. Lidars and hd-radars and new types of fusion sensors will be so cheap you'll be stupid not to add it even for just nighttime or fog to add reliability/safety.

Plenty of engineering and advanced sensing will bridge the gap until, if ever, we figure out how to implement real intelligence in machines and train models with less than the equivalent of the energy consumption of a small town.

 

[mborkow]

I feel like those arguing that Tesla is ahead of Waymo haven’t spent a lot of time in a vision-only Tesla with FSD (supervised or beta). The angle of the sun on my drive this AM caused the car to give me an alert that the system was degraded (because the sun was right in its eye) and there were no visualizations (the road was still there but it couldn’t identify anything other than the road…no cars, bikes, pedestrians, etc).

 

[HitchHiker71]

I feel like those arguing that Tesla is ahead of Waymo haven’t spent a lot of time in a vision-only Tesla with FSD (supervised or beta). The angle of the sun on my drive this AM caused the car to give me an alert that the system was degraded (because the sun was right in its eye) and there were no visualizations (the road was still there but it couldn’t identify anything other than the road…no cars, bikes, pedestrians, etc).

I agree. FSD 12.3.3 still has issues in any kind of inclement weather - especially heavy rain/snow for example - we experienced a sudden downpour using 12.3.3 the other day and the system disengaged itself after slowing down several times to the point where I would have had to disengage it manually anyways because it was going slower than surrounding traffic. This is and always has been an inherent limitation with the Vision Only system IME. It's admittedly better in moderate rain than 11.4.9 was - but it still has a ways to go IMHO.

 

[Dewg]

Again, ponder on why not radiology is unsupervised yet. Still images. Not time critical. A simple CV-application compared to driving, right?

I'll refer you to autonomous robotic surgery - we're getting there.

 

[Joey D]

Doesn't matter how much compute you have if you don't have the data or the techniques to explain or validate the system. Again, data curation and ensuring regressions is the hard part for safety critical systems in ML. And, in the robotaixs that are available today (Waymo, Cruise, Zoox, Baidu, PonyAI) have spent the last 8-10 years on getting to driverless by carefully adding nines by all means possible.

Again, ponder on why not radiology is unsupervised yet. Still images. Not time critical. A simple CV-application compared to driving, right?

Sure. Good luck with that. I'll go out on a limb and say that no one will deploy driverless vehicles with CV/ML only this decade, and most likely not in the next either unles there are some seriois research breakthroughs.

For sensing, Lower hardware costs beats the sh!t out of ML scaling attempts where you need exponential training data sets for linear progress. Lidars and hd-radars and new types of fusion sensors will be so cheap you'll be stupid not to add it even for just nighttime or fog to add reliability/safety.

Plenty of engineering and advanced sensing will bridge the gap until, if ever, we figure out how to implement real intelligence in machines and train models with less than the equivalent of the energy consumption of a small town.

Your argument seems to be "Machine learning is hard."

No *sugar*.

Waymo et. al have limited level 4/5 systems yes. Yet Tesla has orders of magnitudes more miles driven / monitored. Again, no one is arguing against how difficult a problem it is that is trying to be solved. The argument about data gathering, pipeline, curation, etc....that bodes in Tesla's favor from where I stand.

And again, I am not saying that Tesla will solve this tomorrow or even next year. What I am saying is that the evidence I have seen is that Tesla is improving dramatically, and that I see a path to them solving it: generalized level 3 autonomy for sure, and even level 4/5,

I do not see Waymo on the same path. They don't appear to be even trying...they appear to be content with "intra-city" type robotaxis with expensive on-board hardware (compute and sensors.) This is not to say that Waymo's approach is not valid for what they appear to be after...but it is a different goal than Tesla.

 

[Joey D]

I feel like those arguing that Tesla is ahead of Waymo haven’t spent a lot of time in a vision-only Tesla with FSD (supervised or beta). The angle of the sun on my drive this AM caused the car to give me an alert that the system was degraded (because the sun was right in its eye) and there were no visualizations (the road was still there but it couldn’t identify anything other than the road…no cars, bikes, pedestrians, etc).

It's not a matter of who is "ahead or behind." It's what goal each is trying to achieve.

Waymo is "ahead" of Tesla in the autonomous "intra-city" robotaxi implementation. (Though I believe they still have a fall-back ability of remote operation in low confidence situations? So not sure on this.)

Tesla is ahead of Waymo for a "general use, (wide scale deployable) self-driving solution."

BTW...I have spent quite a bit of time in my vision only Telsa with FSD Supervised.

 

[spacecoin]

Your argument seems to be "Machine learning is hard."

No. I'm saying safety critical applications and ML don't mix without massive amounts of engineering, which you seem to brush off as "crutches".

Waymo et. al have limited level 4/5 systems yes. Yet Tesla has orders of magnitudes more miles driven / monitored. Again, no one is arguing against how difficult a problem it is that is trying to be solved. The argument about data gathering, pipeline, curation, etc....that bodes in Tesla's favor from where I stand.

And again, I am not saying that Tesla will solve this tomorrow or even next year. What I am saying is that the evidence I have seen is that Tesla is improving dramatically, and that I see a path to them solving it: generalized level 3 autonomy for sure, and even level 4/5,

Tesla has convinced you that there is a magic path to the fairly land of full autonomy. I get it.

Newsflash: Not. Going. To. Happen.

Tesla will likely deploy first in the LVCC tunnels if they ever get to driverless, then in a sunny warm town, et.c. et.c like everyone else. There are no shortcuts.

 

[phxazcraig]

The Johnny Cab, Seems to me that RoboTaxis will be for specific routes i.e. Airport Terminals to nearby Hotels Rental cars other transport. It would be a logical first step, known routes repetitive, where other vehicles would fill in while others are charging.

Now this is where I could see a Model Y turned into a robotaxi - hotel maintains a MY to pick up guests at the airport and get them back to the hotel. A very designated route done over and over. Still not sure how someone opens the doors, and it wouldn't work for a handicapped rider with luggage.

 

[Joey D]

No. I'm saying safety critical applications and ML don't mix without massive amounts of engineering, which you seem to brush off as "crutches".

I am not brushing anything off. I am saying they are not insurmoutable. You seem to brush off as "impossible, unless you do it the Waymo way."

Tesla has convinced you that there is a magic path to the fairly land of full autonomy. I get it.

No, you don't get it. No magic. It will take lots of data, compute, and engineering. (Which admittedly, might appear as "magic" to the layman.)

Newsflash: Not. Going. To. Happen.

Like I said, you seem to dismiss Tesla's approach as "impossible." I say, possible if not probable, given the evidence we've seen so far.

 

[Joey D]

There are no shortcuts.

And you think that Tesla's approach...billions and billions of dollars, massive AI compute, intricate data pipeline, etc, a billion driven miles recorded , "zillions" of bytes o video storage, simulation, their own AI chip design for both TRAINING and INFERENCE..... etc.......that is what you call a "short cut?"

Newsflash: it is DIFFERENT than Waymo's approach.

Different approaches are just that: different. I could argue that Waymo is taking "short cuts" by using Lidar and massive compute "on vehicle", etc., rather than do the heavy lifting that Tesla is doing.

But I would not say that, because I don't have a one-track mind. Again, Waymo's approach is different. Not inherently better or worse...it depends on the end goal. Why is this so hard to understand?

 

[Daniel in SD]

And you think that Tesla's approach...billions and billions of dollars, massive AI compute, intricate data pipeline, etc, a billion driven miles recorded , "zillions" of bytes o video storage, simulation, their own AI chip design for both TRAINING and INFERENCE..... etc.......that is what you call a "short cut?"

Newsflash: it is DIFFERENT than Waymo's approach.

Different approaches are just that: different. I could argue that Waymo is taking "short cuts" by using Lidar and massive compute "on vehicle", etc., rather than do the heavy lifting that Tesla is doing.

But I would not say that, because I don't have a one-track mind. Again, Waymo's approach is different. Not inherently better or worse...it depends on the end goal. Why is this so hard to understand?

There’s no reason that Waymo couldn’t eventually scale to cover the entire US. They’re just taking a traditional approach of making a minimum viable product first and then work to reduce costs and increase scale. Similar to what Tesla did in developing the Roadster and then the Model S and then the Model 3.

 

[Cal1]

I just don't get the eternal Robotaxi optimism. Auto park, summon doesn't even work. I just don't get why anyone would think Tesla will be able to produce a car that can function autonomously. FSD is no where near close to ready for complete unsupervised use. Something a true robotaxi as Elon promoted should be completely able to do. I do want it to work (paid for FSD twice). I doubt seriously that I will ever buy another Tesla but if I do FSD will not be included.

You can split hairs about what does work but overall it just doesn't.

 

[spacecoin]

And you think that Tesla's approach...billions and billions of dollars, massive AI compute, intricate data pipeline, etc, a billion driven miles recorded , "zillions" of bytes o video storage, simulation, their own AI chip design for both TRAINING and INFERENCE..... etc.......that is what you call a "short cut?"

In a word, yes.

Newsflash: it is DIFFERENT than Waymo's approach.

Not where it matters. ML perception and planning in both.

Different approaches are just that: different. I could argue that Waymo is taking "short cuts" by using Lidar and massive compute "on vehicle", etc., rather than do the heavy lifting that Tesla is doing.

What heavy lifting? IKEA hardware for one of the hardest problem in our generation. That's not clever. It's stupid. You seem to think that Tesla has more and better actual training data than Waymo. I'd love to see a source on that. It doesn't show in driving performance. Shouldn't it by now?

But I would not say that, because I don't have a one-track mind. Again, Waymo's approach is different. Not inherently better or worse...it depends on the end goal. Why is this so hard to understand?

I think our different points of view may be rooted in that I am an ML/AI sceptic when it comes to these types of safety critical systems, whereas you seem to believe that ML by itself with solve self-driving.

ML alone is great for research, like drug discovery, finding new battery chemistries etc. It's not currently great for consistency and safety guarantees. In my view you need throw all tools you have to get close to solve this problem: geofence, mapping, hi-res sensing rule based safety rails around the ML etc.etc.

I find it unlikely that in the short term (this decade) Tesla's (or Wayve's) current "pure ML" approach to self driving will succeed for these reasons.

A couple of major breakthroughs in ML might change that. Until then I remain a sceptic, especially when it come to "the boy who cried robotaxi for a decade".

 

[Joey D]

There’s no reason that Waymo couldn’t eventually scale to cover the entire US. They’re just taking a traditional approach of making a minimum viable product first and then work to reduce costs and increase scale. Similar to what Tesla did in developing the Roadster and then the Model S and then the Model 3.

Agree and disagree. If Waymo continues to rely on pre-mapped roads, I don't think it's reasonable think that can be scaled to anything more than a handful of cities. It's not just the initial mapping to worry about, but the constant change (maintenance) that has to be dealt with.

If at some point Waymo is able to ditch the pre-mapping pre-requisite? Sure.

 

[Joey D]

I think our different points of view may be rooted in that I am an ML/AI sceptic when it comes to these types of safety critical systems, whereas you seem to believe that ML by itself with solve self-driving.

Agree and disagree. I am also a ML/AI skeptic. HOWEVER, this does not mean "impossible."

I also don't think (not sure what you think on this) that autonomous driving needs to be 100% safe before it can be approved for use. That is impossible. It will need to be demonstrated to be much more safer than human drivers.

A couple of major breakthroughs in ML might change that. Until then I remain a sceptic, especially when it come to "the boy who cried robotaxi for a decade".

I understand where you are coming from. I was probably as big a skeptic as you in the ability of AI to solve autonomy...right up until I experienced FSD 12 and how it compared to my experience with version 11.

Now I am "cautiously optimistic." I am not saying "it's solved!" nor am I saying "it will be solved in 12 months!"

But I can now see a path. I can see for the first time that Tesla is just *beginning* the process of "chasing the 9s"...where even with FSD 11...I saw them not there yet. Whether my cautious optimism get more "cautious" or more "optimisitc" over the coming months will depend on new FSD releases and what kind of progress is made on the most widely reported issues and frequency of disengagments.

 

[spacecoin]

Agree and disagree. If Waymo continues to rely on pre-mapped roads, I don't think it's reasonable to scale. It's not just the initial mapping to worry about, but the constant change (maintenance) that has to be dealt with.

If at some point Waymo is able to ditch the pre-mapping pre-requisite? Sure.

It's easy to do map upkeep if you have the sensors needed on the fleet. In general, maps are a prior (like humans driver better where they've driven before), and it let's you see occluded things like traffic signs and let's you see what's ahead (even if you can't see it yet physically) . Someone called it a "time-sensor".

Waymo have said over and over again that they don't rely on that the map is true in order to drive. That should be obvious if you see how they navigate in and around roadwork today.

 

[spacecoin]

I was probably as big a skeptic as you in the ability of AI to solve autonomy...right up until I experienced FSD 12 and how it compared to my experience with version 11.

I had this experience in 2009-2011 when I saw about the same performance from Waymo as FSDb has now...

We'll see what trajectory v12 and beyond takes us on over the coming months.

 

[Daniel in SD]

Agree and disagree. If Waymo continues to rely on pre-mapped roads, I don't think it's reasonable think that can be scaled to anything more than a handful of cities. It's not just the initial mapping to worry about, but the constant change (maintenance) that has to be dealt with.

If at some point Waymo is able to ditch the pre-mapping pre-requisite? Sure.

How much does it cost per mile to create maps? Couldn’t AI also be used to lower that cost?
I wonder if humans drive better in areas where they have a mental map? I feel like I do.