Elon Musk claims Tesla Robotaxis are now driving without safety monitors in Austin
 |
| Hey, why is that black Tesla following us? Credit: Tesla |
Tesla now operates Robotaxis in Austin, Texas without human safety monitors in the driver's seat.
While the above sentence is technically correct, the company may have just moved the safety monitors to another car that follows the first one.
Tesla CEO Elon Musk announced the news on X on Thursday. "Just started Tesla Robotaxi drives in Austin with no safety monitor in the car," he wrote. The official Tesla account on X went a step further, calling the ride "full unsupervised."
SEE ALSO:Elon Musk: Tesla FSD will soon become subscription-only
One "unsupervised" Robotaxi ride in a Tesla Model Y is detailed in a video by Joe Tegtmeyer, retweeted by Musk and Tesla. At one point, we see that a black Tesla Model Y is following the Robotaxi. "You'll also notice behind the Model Y is a chase car, I think they're using that for validation," Tegtmeyer says. The black car keeps following the Robotaxi throughout the ride.
Neither Musk nor Tesla have explained what the second car is for, so it's hard to make any definitive statement about it. It is, however, very possible that the car carries humans who are ready to intervene if something goes wrong with the Robotaxi ride, which sounds a lot like safety monitors.
We'll see whether Robotaxis in Austin will start making autonomous rides without another Tesla following them. For now, this could just be a cool trick that allows Musk to brag about Robotaxis making autonomous rides with "no safety monitor in the car."
Meanwhile, Tesla's chief competitor Waymo is carrying out fully autonomous rides in six major cities in the U.S., having added Miami to the list just days ago.
Topics Self-Driving Cars Tesla
Gee whiz. Do you think Musk is overly optimistic. I am glad I don't own Tesla stock
ReplyDeleteMy Tesla already drives itself to anywhere I want to go. They probably have monitors to keep the lunatics from vandalizing the cars.
ReplyDeleteAssuming that's true (it's not), wouldn't that mean that all robotaxis in the future will need to have a monitor riding along? If Tesla is concerned about vandalism in a highly controlled environment now (they're not), think it would be any different in downtown New York, Detroit, or Chicago? And if they need to pay a monitor to be in the car, that eliminates the financial advantage of an autonomous taxi.
DeleteThe reality is, the monitor is definitely in place because the autonomous software is not ready to be unsupervised. But it's funny how this attempt to explain away this need would actually be much worse for Tesla in the long-term.'''
Sounds expensive to me.
ReplyDeleteTesla is going to be a great case study someday in the advantages and disadvantages of centralized decision-making. Musk is clearly brilliant at some things, and was able to get some good results, but because he has so much control, and so little filter, his bad decisions will drag the company down with him.
Rivian seems to be doing a better job of keeping the main thing the main thing and whenever they come up with something new that would be a perfectly good main thing on its own (which is common in the EV startup space), they spin it off into another company that can make their own thing the main thing.
There's nothing at all wrong with pursuing self-driving taxis as a business model. But do it in a subsidiary while letting the main company focus on developing a profitable Model 2, making a cybertruck that doesn't suck, updating the S and X to be competitive at their price ranges, making a true halo roadster 2, and so on. One at a time if that's what their limited engineering budget allows.
Surely the stock price could still be manipulated to their satisfaction.
You do realize that Rivian is on life support and near death - while Tesla is at $1.3 Trillion market cap and so to be the worlds biggest most profitable company? Don't let facts get in the way of your brilliant observations. You'll be BIG envious when Elon collects his $ Trillion performance bonus - upping his net worth from a meager $800 Billion as it stands now. Come on - tell the truth - you're so smart you voted for Kamala!
DeleteAre you a parody account?
DeleteOpenweb needs a "based in x country" feature like twitter.
Meanwhile, Waymos work pretty well.
ReplyDeleteWorlds slowest car chase
ReplyDeleteThey've got chase cars, currently. https://xcancel.com/JoeTegtmeyer/status/2014424041147310233#m
ReplyDeleteWas hoping this would be a bigger step forward. This is just moving the safety monitor from the passenger seat to behind the car with a killswitch.
Safety first.
DeleteWould not be safer if the safety driver was behind the wheel?
DeleteThis is not exactly rocket science. Obviously Tesla has safety standards, which they are trying to meet while gradually relaxing oversight. If you owned a company, how would you do it? You seem to be great at criticizing.
DeleteIt's the correct move. Inevitability Robotaxis will be under gov scrutiny. Tesla will show that they took every reasonable precaution in terms of safety.
DeleteChase cars are a huge step forward imo. It would be extremely negligent to not have a chase car. They’d get hella shit for it. This is a good way to roll out the new patch.
DeleteThis feels more optics than anything. The safety driver in the passenger seat had very limited controls of the car to begin with. All they had was a stop in lane button or opening the door. The chase car probably has similar controls.
DeleteI don’t understand why there’s a step between “limited control” safety driver and “limited control” chase car other than optics. The true jump is when Tesla moves away from the hawkish monitoring of each individual vehicle.
Elon also said he needed a few more billion miles for safe Unsupervised, so there’s a contradiction here too.
This comment has been removed by a blog administrator.
DeleteThey are now only 5 years behind
DeleteNot really. I’ve seen Waymos completely stunlocked in the middle of intersections in SF for months, causing insane traffic problems.
DeleteI think the issue (at least for me) is the transparency. If indeed they are concerned like you say, then say "we have a chase car" but instead they leave that part out. Report when remote monitoring and control is going on... don't make it look more than it is. This, along with lofty promises that never hit time lines, makes people question everything.
DeleteHow's it any different than waymo remote operators?
DeleteGive me the actual X link.
DeleteI feel like it would be more effective to have a remote video feed to a human. anyone who has followed another car knows the million ways you can lose them.
DeleteSooooo can we get our hardware upgrade now?
ReplyDelete*in the car
ReplyDeleteI would take this taxi over any others for the sole reason of having no tips.
DeleteWeMo has teleoperators too. It would be insane and unsafe for any company to operate a AV fleet without that in-place.
DeleteExactly so saying "any safety monitors" is false.
DeleteThe remote assistant to vehicle ratio for Waymo is on the order of 1 human for several hundred vehicles, if not higher. That’s not really the same as having a safety monitor watch everything the vehicle does. If Tesla has. Similar ratio I think saying it is autonomous without a safety monitor is a fair claim.
DeleteEvidence or bust
DeleteI have no idea what Tesla’s is currently. I’m merely stating that generally the existence of remote monitors doesn’t mean the car isn’t driving autonomously, as it’s basically a customer support position for the vehicles.
DeleteBefore Cruise went bust in 2023 it had a ratio of about 20:1 vehicles to remote operators. Waymo certainly has an even better ratio, and with headcount growth compared to fleet size growth you can estimate that it is at least a 100:1 ratio, but is likely much higher than that.
yeah and i'm sure that ratio has changed over time. they were probably super cautious in the beginning. they've already hired people from Cruise, they probably have ex Waymo people working there advising them.
DeleteOr maybe Tesla just did it better.... Hmmmmm. It was specifically stated that the agent will operate without any external connectivity needed during multiple Tesla AI-day events.
Deletelmao the delusion is strong with this one
DeleteThis comment has been removed by a blog administrator.
DeleteThe question I have is when will this be available for Intercity travel? I want to be able to use robotaxi to travel from Austin to Waco/Dallas.
ReplyDeleteThis comment has been removed by a blog administrator.
DeleteThis comment has been removed by a blog administrator.
ReplyDeleteIt’s probably on 15. Even the video of the model y self delivering to the customer was on a different version. They downgraded it after delivery.
DeleteNo, v15 (assuming the 10x parameter model will be called v15) is coming later, possibly next month. This is the smaller model.
DeleteThe larger model isn't out yet?!?
DeleteVery awesome if true. Our cars will only get better. My other experiences with increasing model params have been surprisingly good. This could be special.
No, it's not out yet. Elon said in December that it would likely be released in February.
DeleteIt does suck, so much hesitation
DeleteThis comment has been removed by a blog administrator.
DeleteHard to see this as my FSD on my Model Y still isn't as smooth as I hope. I guess it's only specific areas like waymo what their doing.
ReplyDeleteWhat version are you on
DeleteHW4 14,2.2.3
DeleteIs this on hw4?
ReplyDeleteit's a juniper so yes
Deleteall robotaxis are hw4
This comment has been removed by a blog administrator.
DeleteDo the ones in San Fran have safety monitors?
ReplyDeleteYes, in the drivers seat.
ReplyDeleteAre they still driving or has Tesla finally gotten the permits?
DeleteThey've never been driving. The car does the driving; they are there to intervene if necessary.
DeleteFrom what I recall, they're driving like standard taxis with the drivers maybe using FSD like normal people do. No different than an Uber who drives a Tesla. They don't have the permits to operate autonomously at least not since last October.
DeleteI literally live in SF and have ridden in robotaxis a dozen times, lol. The monitors are not driving.
DeleteCherry picked routes and a chaperone vehicle following. Nice try Elon.
ReplyDeletePretty wild to see unsupervised FSD for real!
ReplyDeleteI wonder what the differences are between what these Robotaxis are running and the publicly available FSD (Supervised) on Tesla vehicles.
Is it better mapping of the local area? Is there new hardware (cameras, compute, etc.) in the Robotaxis?
Hardware is the same. They are stock refreshed Model Ys.
DeleteThey said no safety monitors, they didn't say no one is watching and controlling the vehicles remotely
DeleteIt’s 100% the mapping
DeleteIt probably is, but not necessarily. It could just be that they have only VALIDATED it to a robust enough safety standard within this strict geofence area. It doesn't mean that if it went outside the area it would be significantly less safe. Whereas with Waymo for instance, if it went outside a mapped area, its functionality WOULD be significantly reduced.
DeleteThe presence of a geofence for both does not mean equal reliance on mapping.
"Whereas with Waymo for instance, if it went outside a mapped area, its functionality WOULD be significantly reduced."
DeleteWhat are you basing this on?
But if they need to manually validate every place in the world for unsupervised, how is it functionally any different than Waymo? They’d still need to drive through every road, every city…
Delete> But if they need to manually validate every place in the world for unsupervised, how is it functionally any different than Waymo? They’d still need to drive through every road, every city…
DeleteTesla has million users who can do 99% of the validation work for them for free. They can automatically enable streets when user supervised FSD goes well, and avoid those with interventions.
That’s too bad google doesn’t have a comparable service, with millions of miles being mapped. Maybe if they did, they could call it Google Maps, and maybe with street level they could call it Street View. And maybe they could have people all around the world driving in cars mapping with cameras and lidar where they drive for Street View. Maybe one day! 🙄
DeleteGoogle maps helps nothing in self driving validation. It's not about mapping, it's about validating how your self driving car handle this street. There is no other way than just do a test drive with self driving on this street. Tesla has users that do it for free, others have to perform all testing on their own.
DeleteYou know Teslas use Google Maps for navigation, right?
DeleteTesla is just slowly realizing that they need to do what competitors are doing. Like how Elon claimed that if an autonomous car needs geofencing (Waymo’s rolling out by city/area), it’s not truly autonomous… and that’s what they’re doing now.
DeleteThey’ll eventually realize they need more sensors and LiDar, but they problem is that they would then need to retrofit all sold Tesla with that new hardware, since their promise to customers at the time of sale was that all Teslas have the hardware needed for self-driving capability 😆.
Probably why it's strictly geo fenced at the moment
DeleteThe mapping is the same also. The inference model weights are different.
DeleteWhy would you assume it's a different model?
DeleteBecause it was mentioned in one of the interview. They run a model that is trained extensively in the area of Robotaxi services.
DeletePretty sure they never said they're using a different model, outside of the period of time when public FSD hadn't been updated in several months.
DeleteHomogeneous model, but newer weight configurations, more parameters, and reasoning logic. Basically, it's just a newer version of the software.
DeleteYou just made all that up. Also, the model largely is the weights. A model with different weights is basically an entirely different model.
DeleteBelieve what you want. It's simple. When inference training is done, the inference model weights are configured. This happens in every version.
DeleteCorrect, which produces a new model. But there's no evidence they're using a new model here.
DeleteChanging the weight values in the model doesn't change the model. Changing how the model is trained changes the model.
DeleteThe training is literally what sets the weights. You seem to think the weights are just a few dials that the devs turn to tweak the model behavior. No, the model is billions of dials, and the dials are turned via training. These dials are the weights. They're by far the most significant part of the model.
DeleteNo, I don't think that. Training is now mostly autonomous. I'm trying to be helpful in your understanding.
DeleteModels are defined by how many parameters they have. Models, especially newer ones can have parameters that have yet to be explicitly trained to achieve certain confidence thresholds. Inference training is continual and dynamic. The fleet is specifically queried for videos that are relative to parameters for the current inference model that need more training. As these videos are consumed, interpreted and auto-labeled (no longer manual) distillation into inference weights are done, for the current model. Once certain confidence thresholds are reached across all parameters, a new weight configuration for the current model is produced. The weights are simply stored as separate values in a large text file and sent out to the fleet as a new minor FSD version. When the model's weights are changed from continuous training, this does not change the model's parameters, it only defines the existing parameters behavior.
Models change when new parameters are defined, but not necessarily yet fully trained. These parameters are also technically autonomously established and only numerically limited. This reduces the possibility of significant regressions from too many new parameters that don't yet have explicit matured training or inference. Typically these newer parameters outputs are ignored but observed which has been referred to as "shadow mode" in the past. In subsequent versions of the same model, it will receive newer weight configurations, but not new parameters, meaning the model stays the same. When the weights are configured, the model is again observed, if the confidence level of the new parameters reaches a certain threshold, then the new parameters can be taken out of shadow mode. Not all parameters need to be taken out of shadow mode at the same time and it can be done in batches, all for the same model. This should give you a hint of how version numbers are internally determined.
Basically: Models do not change until new parameters are defined, those parameters are constantly configured with weights within the same model, and parameters can lay productively dormant but developmentally observed.
Wait hold on, you think they're still labelling? What specifically do you think they're labelling, and for what purpose?
DeleteThere is no they, only it. Labeling still occurs, but autonomously. Labels are not defined by humans and the training system determines its own labels. It's best to think of labels now more as conscious observations. It doesn't know the word dog, but it has seen it before, which is still, technically, a label.
DeleteNope, labels are example outputs. The only outputs now are driving controls (acceleration, steering, turn signals, gear shifts), which are provided by Tesla customers driving the cars.
DeleteYou don't understand this system. You're thinking of the pre-v12 architecture, combined with nonsense.
I'm only divulging this info as an exercise. I get why you'd think that way, but you really shouldn't be so certain and dismissive to information that challenges your understanding. Auto-labeling is necessary, autonomous and integral to the training system. Yes, there are inputs and outputs, yes driving controls are considered in labeling, however ALL observations are considered in labeling, including temporal ones. This is the magic that auto-labeling unlocked, the labeling of heuristics, which is how an observation's behavior is observed over time. This includes anything that moves, this can also include, audio! It's mind boggling what auto-labeling unlocked and I still get excited over it.
DeleteI don't know why you still think there's labeling. This is an end-to-end neural network. Do you actually think objects like vehicles are being labeled (auto-labeled or otherwise)?
DeleteLabeling is what helps define new parameters for the inference models based on the training system's perceived importance of each observation. Even though humans are not involved, labeling still happens within the training system.
DeleteI'll ask you again: What do you think is being labeled? Vehicles?
DeleteNo sir, it's quite literally everything. The way the auto-labeler works it would be impossible for any amount of humans to replicate. The main thing I think you're misunderstanding here is that labeling happens at the large training computer and is later distilled into inference parameters.
Delete-Labeling is about the training signal (what the training network should match).
-End-to-end is about the inference graph (pixels -> driving output)
"Everything" includes vehicles. If you think that vehicles are labeled, you're already wrong. That's pre-v12.
DeleteYes, labeling produces example outputs, which the network learns to imitate via training. But vehicles are not one of the outputs of an end-to-end driving neural network. That's what you don't understand.
I think you’re still mixing up training with driving.
DeleteWhen the car is driving, it can be “end-to-end” (camera pixels → steering/brake). No “vehicle boxes” need to be passed around as inputs. But when Tesla is training the system, they still need “answer keys” so the network can learn. That’s what labeling is. Most of those “answer keys” are now made automatically from the video (auto-labeling / reconstruction), not hand-drawn by humans.
So: labels still exist in training, they just aren’t necessarily part of the live driving model’s inputs/outputs.
It's worth noting for now they have a chaperone vehicle that follows behind.
Deletehttps://x.com/JoeTegtmeyer/status/2014410572226322794
This comment has been removed by a blog administrator.
Deletehttps://imgur.com/a/sCaNAYE
DeleteNope. Just cherry picked, simple routes and plenty of pressure from Elon to show something to the public.
ReplyDeleteIt was obvious 5 years ago that Tesla was going to win the SDCF race. It’s only become more obvious.
Uber self driving shut down.
Cruise shuttered. (Buddy who was high up at cruise says they didn’t even follow Tesla autonomous because they did not consider Tesla serious competition)
Waymo is next. Frankly I’m surprised they are still devoting resources to it. Probably trying to time the inevitable write down.
And it has always been obvious this was going to happen.
I really doubt Waymo is vanishing anytime soon considering how rapidly they’re expanding and where they’re expanding. They’re not just in the US anymore.
Deletewaymo does not have a path to beating Tesla.
DeleteSelf driving fleets are winner take all
In order to win that race, Waymo needs to build approximately 1 million cars ($200B), and hire enough staff and build enough facilities to maintain a 1 million car fleet. Call it 175,000 people, based on 1 hr cleaning and maint per car per day. And maybe 100-200 facilities.
Conversely-
In order for Tesla to win that race, Tesla needs to roll out an over the air update.
So for Waymo to have a realistic chance of winning, Google needs to essentially bet the entire company on Waymo. And It will be a multi year project to hire that many people, build that many facilities, and build at least 1 million cars. And at any time Tesla could roll out an OTA update and take the crown.
Google is not going to bet everything on that. Google knows they would do much better putting that capital investment into regular software AI development, which has much higher returns and much lower risk. And it is their core competency.
Building and operating a fleet of millions of cars is not a Google core competency.
Google will eventually bow out, say that all of the AI research from Waymo is getting folded into their normal AI work, and that while it was ultimately unsuccessful it was worth it for the AI progress.
I would bet a million dollars on it.
Ok, very serious question: is Waymo still losing billions per quarter? Have they found any path that remotely leads to profitability?
DeleteObvious how? There is literally 0 things to suggest Tesla will be the leader in self driving cars other than Elon just making things up and saying they’ll happen in 6 months.
Delete
Deletehttps://old.reddit.com/r/teslamotors/comments/1qk1vak/robotaxi_rides_without_any_safety_monitors_are/o16bv89/
Just to avoid re hashing the same statements..
Basically Tesla is too close and Waymo is too far. Waymo won’t risk the capital needed to stay in the race.
And Google doesn’t want to own and operate a fleet of millions of cars. Google is making shit tons of money in software.
Tesla doesn’t own a single one of those cars that you’re claiming will be active in an over the air update.
DeleteYou do realize if Tesla wants a fleet of actual robotaxis they will need produce a fleet as well, for the tune of tens of billions as well..
Correct. That’s why teslas robotaxi model is brilliant. The customers pay Tesla for the car, and Tesla gets a cut of the rideshare.
DeleteAnd Tesla has the option to divert excess factory capacity to building their own fleet, over time.
Waymo has neither of those options.
Waymo could contract manufacture cars and then resell them to “franchisees”. But they really can’t. Waymo cars with the tech package are too expensive ($180k ish) to generate a meaningful profit when financed and with Waymo taking a cut.
Basically the Waymo model is obscenely capital intensive. The Tesla model is capital flowing. Every additional Waymo costs Google say $250k in hardware and overhead with a long payback period, every additional Tesla on the road generates Tesla $50k on day one. Only one of those situations is infinitely scalable.
Do you have a source for “Waymo cars w tech pkg are… $180k ish”?
DeleteGoogle search for “how much does a Waymo car cost”
DeleteA Waymo car's total hardware cost, including the vehicle and advanced self-driving tech (sensors, computers), is estimated to be around $150,000 to $200,000 for current models.
Got it. So today’s cost is closer to $70k … and dropping significantly going forward with scale.
DeleteSo basically, your argument and cost projections are completely wrong.
Todays cost is $180k, until it goes down. So basically, your argument and cost projections are completely wrong.
DeleteThe critics and shorts have used the "he's just making it up" line for over a decade. This line of argumentation has not worked out for them. But I guess you feel like you are different.
DeleteIf you tell me something is coming in 6 months and it takes 8 months, you were just wrong.
DeleteIf you tell me something is something is going to take 6 months and I still don’t have it after 10 years, you were lying.
It’s funny you’re responding with such smugness when you haven’t provided even a shred of evidence to support your thesis, and all Tesla has done is fail to deliver thus far.
Given that the evidence is right in your face, I really doubt anything I can say will reach you. What you are interpreting as "smugness" is just experience effectively countering slogans.
DeleteWhat I love is that you think you are the first person in the world to have cracked "Elon-Time". Well thanks. But we already know about his tendency to underestimate timelines.
The evidence of what? That Tesla has literally been making things up for years?
DeleteI’m not talking about optimistic timelines, delayed products, or prices changing between announcement and delivery. Every manufacturer does that, and none of it is “making things up.”
DeleteIf you’re alleging fraud or fabricated data, say so and point to evidence. Otherwise this is just hindsight criticism of product roadmaps.
Although I just suspect you are going to throw out some more "Tesla Shorts Greatest Hits". But hey, maybe your previous posts do not reflect your actual thought process.
Feature complete full self driving by the end of 2016.
DeletePaint it black video released same year.
Proven in court to be fabricated.
It was a lie, it was literal fraud.
You’re conflating three different things and calling it “literal fraud,” which is not what any court found.
DeleteThe 2016 “feature complete” claim and the demo video fall under marketing and product roadmap claims. Courts that looked at this did not rule that Tesla fabricated capabilities or falsified data. They ruled that a promotional video could mislead viewers and that disclaimers mattered. That is a consumer protection issue, not a finding of fraud.
Fraud has a specific legal meaning. It requires knowingly false statements of fact made to induce reliance. Missed timelines, optimistic projections, or a staged demo do not meet that standard, and no judgment says they do.
If you are claiming Tesla committed fraud, point to the specific ruling where a court found Tesla knowingly falsified capabilities or data. Otherwise this is criticism of marketing being mislabeled as fabrication.
The definition of fraud:
Delete“wrongful or criminal deception intended to result in financial or personal gain”
“a person or thing intended to deceive others, typically by unjustifiably claiming or being credited with accomplishments or qualities”
Elon knew the paint it black video was fake, it was not what they were advertising, it was not full self driving. They edited the video and cut out a crash at the direction of the CEO of the company.
The same guy who released this on stage to support his increasing valuation and show progress of FSD.
That is literally fraud, or at the least material misrepresentation to investors. A known deception.
That is Theranos level stuff, so is the fake solar roof demo, arguably the roadster that never came out.