The auto manufacturer approach is also showing progress. In CA and NV you can buy and operate a Mercedes with Drive Pilot, which is Level 3 certified. In the right (very restrictive conditions which essentially come down to "sitting in highway traffic on your commute") you legally do not have to pay attention to the road and can read/watch/work/etc.
This is about the peak of what you can get with automated lane keeping and braking. I don't see any route from this point to anything like level 4.
> This is about the peak of what you can get with automated lane keeping and braking.
Are you saying that within 5y, say, we won't see a level three system that's able to handle full highway speeds?
well yeah, that is the definition of level 3. That's not going to change.
They are limited by:
> Our technology relies on a digital high-definition map that provides a 3D image of the road and the surroundings with information on road geometry, route characteristics, traffic signs and special traffic events
IMHO Tesla's "F--- it all just use NN" approach does get around the pre-generated 3D map requirement. Even if it is not much more than Level 3.5 at the moment. Pretty funny to see it accelerate to 65mph in a parking lot b/c it thought it was on the nearby highway though.
It's not even level 3, and couldn't be certified as such. It's not reliable enough for the person in the driver's seat to focus their full attention elsewhere.
>> level 3 is still basically lane keeping and auto cruise control
> well yeah, that is the definition of level 3
No, it's the definition of level 2: https://www.sae.org/blog/sae-j3016-update ("example features: lane centering and adaptive cruise control at the same time")
Am I missing something? Lane keeping, cruise control and offset keeping is level 3. That's what this feature offers according to a few posts up.
It makes no difference to the conversation. I'll stop poking the bear here.
I'd personally never trust an autopilot unless it's either backed by human-level AI which has also had years of driving experience, or it's in some very highly constrained environment (maybe airport bus going from gate to plane). Out on a highway or public road system is the most unpredictable environment possible.
> All it takes is for something non-routine to happen such as a car ahead reacting to an animal, or swerving as the driver reaches for something or spills coffee on themself, or a wheel come off a car (I've seen it happen to a car in front of me), or a car crosses the center median in opposite direction (which left my ex-boss hospitalized for 6 months)
Inherent in this statement is the assumption that in such types of events, a human would necessarily do better than the machine. Each of these are extremes and I doubt that most human drivers would be able to react to avoid an accident or damage most of the time.However, the human has the major advantage of having a brain and being able to understand the consequences and potential outcomes of something in real-time as it is unfolding. I doubt most autopilots would understand the situations I mentioned - certainly not unless they were specifically pre-programmed/trained into the system. Would an autopilot even see what is going on inside a car if a driver is bending down below dashboard, or fighting with passenger, for example?
Your average driver is already at the limits of human capabilities with no room for improvement unless they go train to drive like F1 racers. So comparing humans to the current state of tech seems silly given that the tech will assuredly one day be faster than a human.
I'll be happy to take it seriously after that happens. In the meantime, I'm sticking with my belief that unreliable automation is worse than no automation.
You just dreamed up this statement and immediately believed it.
You cannot train a F1 driver for highly dynamic events at 20k rpm and 4000 NM forces at the axles, you need automated controllers for that. You can train him for simple things, like gear switching (5ms) and hitting the breakpoints right. But an AI will be at least 10.000 better than a human on this.
You need a slow human brain for the stupid mistakes instead.
This is nowhere near true. The skill range from average driver (who knows nothing about car control or reacting correctly) to F1 drivers is vast. There's a huge gray area between these extremes that would be improve safety is people chose (or were forced to) get some training.
Even something as basic as taking a car control clinic once a year would improve the average driver skill and safety by a huge amount.
If we had the same lawyers available that Tesla does, humans too would be not responsible for much of anything.
I merged hard right before it was clear. I assumed the car on my right would do the same, and the driver further behind would break in time. They either made the same assumptions as me or took my queues and acted accordingly and everything turned out ok.
It is pretty amazing how people can coordinate on-the-fly.
Sadly its self-ownership is only "according to legend" rather than anything battle-tested.
How do you plan to do that? Will you wrestle the code away from Waymo? Or do you plan to put in the long years of thousands of man hours to develop it and all the costs of the hardware while you do it?
edit: What to do when people get brain chips?
There would be a legal firm to handle direct tasks, and vetting contractors, then an auditing legal firm. A CEO-type, and various checks/fail-safes.
Periodically hire competent groups to come up with new ideas and have a policy for updating the policy.
One of the "ideas" people would eventually realize the need for failsafes for the failsafes: taking control of the entity's assets would become extremely profitable. So you need some kind of secret society to keep an eye on things and execute the assassinate-and-reset plan when necessary.
Naturally its influence would need to turn malign, after it was too powerful for anyone to stop it…
A friend from Türkiye told me in his village they take turns taking everyone's cows to pasture every morning and back home at the end of the day. Some homes are abandoned, people died, the gates are open, the cows still living there just join the heard and go back home at the end of the day.
Horses or dogs should be able to learn to walk from A to B if they get food there. If you don't pay them they should eventually refuse (specially if it is you again) and require some payment in advance. Younglings can be tied to the cart. I think they will know their route eventually.
You could even have a Musk like figure promising it will work next year for the first few decades.
Nor is it possible in the West for two companies to own 100% of each other's shares (i.e., circular ownership). In most of the West, a subsidiary acquiring its parent's stock is treated as a redemption of those shares (i.e., as if those shares were returned to the parent), assuming the transaction is even allowed.
I can easily imagine the same (unprincipled) dynamic applying to an AI lawyer.
IBM had an AI (Ross) pass the bar years ago...I believe it was actually 'hired' as an attorney at a firm in London for routine paperwork..
0) https://www.linkedin.com/pulse/ibms-artificially-intelligent...
1) https://www.dailymail.co.uk/sciencetech/article-3589795/Your...
Won't it be great once we have fully self-driving cars? Heck, I could buy a car and then rent it out to other people like a taxi when I'm not using it, and it would pay for itself. Maybe I could even make a profit!
...
If I could make more money than the car costs to purchase and maintain, without any additional work on my part, why would the company sell me the car at that price in the first place rather than just running the taxi service themselves and keeping all of that extra profit?
https://www.teslafsdtracker.com puts miles to disengagement at 30 and miles to critical disengagement at 300 for all v12.x.y versions. Note: this is crowdsourced data and the users themselves get to decide what's critical and what's not.
As far numbers required to make it fully self driving, it's at least 3 orders of magnitude worse than the big players. Waymo and Cruise routinely had 30,000 miles per disengagement during their California testing. That's one disengagement for roughly 3 years of driving.
On top of all that, Waymo does intensive 3d mapping for their service locations. These maps have to be maintained. Then the cars need sensors that take advantage of those 3D maps. If that combination of intensive mapping and LIDAR sensing turns out to be necessary to get beyond FSD's current and near future performance, then Tesla isn't even at the starting line.
Well, being as how people get to pick and choose when they use it, and that the driver has to remain vigilant at all times, I'm not surprised.
But this is easy to test: stick random people in the car and go to random locations with FSD, see how it works. Why haven't they demonstrated this yet?
I am very skeptical of the "weeks without intervention". It's cool technology, but I never had a single trip where I didn't need to intervene at least once.
It would regularly blow through school zones, failing to read the posted sign.
On a couple of occasions it veered off the road on to the shoulder.
My thinking is the car will never be level 4. It doesn't have sufficient sensors or NN compute power.
https://www.forbes.com/sites/cyrusfarivar/2023/12/04/judge-a...
Not sure if you count this as "legit" or not, but I haven't seen similar incidents from Waymo. (Perhaps I've just missed them - if so, links welcome!)
"Lying to judges" (do you mean withholding material information from regulators?) is not something I'm aware of Waymo doing. (Again, links welcome -- and remember Waymo is not Google.) Nor is it a binary thing. It's one thing to cover up e.g. anti-competitive behavior in the free market, but quite another thing to cover up how you might've actually killed a person on the street.
I have a comma.ai in our minivan and it works great. Much better than Honda's built in lane following tech
The question of Camera vs LIDAR+Camera is a narrow technical question about how to construct a 3D scene. That's it. It says nothing about making sense of this 3D world for which you you have a 3D point cloud and it says nothing about how to actually navigate that world. Say you're driving down the road and there's a bit of construction, there's a guy holding SLOW/STOP sign directing traffic. LIDAR will tell you it's a hexagonal sign, but it can't tell you what it says, you need a camera to read the sign and tell you what it says. It doesn't tell you how to drive, how fast you should go, how much space to give the guy with the sign etc. Everything AV-related which is not constructing a 3D scene is actually the same across all AV stacks, which includes the hardest part - the actual driving itself.
Your example of needing to read a stop sign isn't a great example. At least in North America, a hexagonal sign is always a stop sign. A better example of your point would be a speed limit sign.
the back of a stop sign is often a slow sign, or a do not enter sign. The same shape of sign, but different meaning from different directions. The slow variant is often held by construction workers, hence the GPs example.
> Say you're driving down the road and there's a bit of construction, there's a guy holding SLOW/STOP sign directing traffic.
Here's a picture of a guy holding a hexagonal SLOW sign. They are very common. https://nj1015.com/how-slow-should-you-go-through-constructi...
LIDAR cannot tell you what's on that sign. It cannot read any sign, nor any road markings, nor streetlights, blinkers, stop lights, etc. If you believe that cameras are not capable of being used as input due to dynamic range or any other reason, then that's fine, you just believe that self-driving is impossible (lidar or not). But to believe that one can safely drive using nothing but a completely blank and unlabeled 3D scene (ie, LIDAR-only)? That's pretty crazy.
But Waymo never said you don't need cameras. Hell, they have 29 cameras in each vehicle compared to Tesla's 8.
Your point about their approaches being more alike than different is somewhat true, but you wrongly attribute the LiDAR vs camera debate to Waymo marketing. It's Elon and Tesla fans who started it and incessantly repeat it even to this day. Most rational folks say use whatever you can to get it working (which Waymo did) and optimize later.
So the question is: is LIDAR also necessary or are cameras sufficient? IE, can cameras+motion give you an accurate-enough 3D scene the way LIDAR can. And that's a narrow technical question, and it isn't even the most question when you consider self-driving as a whole.
"LIDAR is necessary" is not exclusively a Waymo talking point - it is shared by all companies using LIDAR, suppliers of LIDAR etc. But it is just a talking point, there's no reason to think it's actually true.
1. Only companies who have LiDAR in their stack have shown fully autonomous driving i.e. driverless.
2. Better 3D scene construction, which LiDAR, unquestionably provides permeates as an advantage throughout the stack.
I do think “LiDAR or not” is a narrow argument. But the advantages are massive and undeniable, so it becomes necessary especially in light of rapidly falling costs.
There are atmospheric conditions and obstructions that lidar can see through that cameras can't.
Cameras also seem prone to being blocked by a small splash of mud/dirt. Is anyone on this thread knowledgeable enough in the domain to know if that's an issue? I thought of it while moving my head sideways to see around a temporary sight obstruction on my windshield. Luckily the windshield is big, and I can move my head. Cameras are small. I guess you just put several so you have an effectively large camera array? It does mean more redundancy is necessary than I would have initially thought.
It's called Level 3 because it is level 3. Mercedes went through an approval process and carries insurance (or a bond iirc, there's a couple options) to comply with California law dictating the use of L3 features. You are legally allowed to stop paying attention under certain conditions and the restrictions to roads or situations is in no way disqualifying, nor is geobounding to only states you are legally allowed to operate in. Also, its available across all of Germany.
The way its actually a marketing gimmick is how few Mercedes has actually made available and the exorbitant cost. They've been allowed to sell in California since June of last year and only have 65 available and 1 sold as of April:
https://fortune.com/2024/04/18/mercedes-self-driving-autonom...
Your mind is going to be blown when you hear about Tesla and the name they give their assisted cruise control.
Tesla already silently abandoned the "just over the AIR one day" approach with a dedicated car announcement.
However the camera+ultra-sonic radars but no lidar is not only Tesla vision, but other companies too.
We don't know what it costs Waymo to operate their car. The fact that they charge money doesn't make them a real business, just as people paying for FSD doesn't make it a real business.
Both are promises until a breakthrough occurs. Waymo is starting small-scale but for a full setup, even if guided by humans here and there. Tesla starts with millions of cars and multiple countries but with far modest functionality.
Waymo is scaling up; Tesla FSD finally starts to look like the promise, with a high chance of a ride with 0 disengagements still on the scale of many countries and launching it also on a different continent right now.
It's interesting to observe how companies with radically different approaches are about to arrive at the same goal almost simultaneously.
https://arstechnica.com/cars/2024/05/on-self-driving-waymo-i...
> "The version of FSD I tried in March [of 2024] was clearly not ready for driverless operation. For example, I had to intervene to prevent the Model X from running over a plastic lane divider, a mistake Waymo would not have made in 2020. So while FSD 12.3 seems superior to Waymo’s technology circa 2018, it’s not as good as Waymo’s technology at the end of 2020."
> During a 45-minute test drive in a Tesla Model X, I had to intervene twice to correct mistakes by the FSD software. In contrast, I rode in driverless Waymo vehicles for more than two hours and didn’t notice a single mistake.
That seems pretty significant.
The only differentiating factor is that Tesla's approach has theoretical benefits, if they can prove it works. If they could hit the same disengagements/mile rate as Waymo, they would likely be ahead since they don't need pre-mapping.
That being said, theoretical benefits are worth about the same as monopoly money. Until they can demonstrate that they can get the same performance without pre-mapping and LIDAR, it's all just conjecture. There are no points awarded for "yeah, well if mine had worked it would have been better!" They might get there eventually, but the current signs don't seem promising. At least not for getting there before Waymo completely eats the market.
Also, at least Waymo's disengagements are known. They're defined by the DMV, it's basically "any time control needs to be taken away from the autopilot". If a human in the car manually takes over, that's a disengagement. If the car decides it can't drive safely and prompts a human to take over, that's a disengagement. If someone has to log in to the car with a joystick to get it somewhere, that's a disengagement.
The only confusing situation I'm aware of is when remote techs give the car waypoints, like if it gets confused in a parking lot. I don't believe that counts as a disengagement because the car is still driving itself, it's basically just failing at pathfinding. That seems reasonable to me, because it's not a safety risk at all, just an annoyance to the business.
Btw, the definition of disengagement is up to the companies, not the DMV https://thelastdriverlicenseholder.com/2024/02/03/2023-disen...
They can't fudge them too hard without the CA DMV yanking their license.
> People are doing 2-3hour drives with FSD 12 with no disengagement. You should check out the recent progress.
2 or 3 hours is impressive in a vacuum, but not all that much compared to the other players. It's an accomplishment for sure, and would have been overall impressive in 2018 or 2020. In 2024, it's only impressive because it can do it without lidar or radar, and even then that may only be because Tesla is the only one doing that so there's really not a comparison point.
It's certainly still an accomplishment, but I'm not convinced that it's economically relevant at this point. Best of luck to them.
To be clear, I dislike Tesla, but I do hope this works out because the alternatives are much too expensive for consumers to actually own. They'll have to use an Uber/Lyft kind of system because the lidar sensors alone cost more than a Lamborghini last I heard.
They do appear to be having basically the exact issues people thought they would have without lidar, though. Namely that different types of sensors are vulnerable to different types of phantom objects, and it's difficult to eliminate those without either having a different sensor that doesn't see those phantoms to sanity check against, or creating issues with failing to detect real objects.
A couple of examples: Video is prone to detecting the stick cyclist on road signs as an actual cyclist, where it's obvious on lidar that it's a road sign. Inversely, lidar can read steam (like that coming out of a sewer manhole) as a physical object where video can tell that it isn't.
It's difficult to use a single signal to fix either of those without having knock on effects that cause failures to detect real objects. It's much easier to use a combination of signals to disintermediate. Cyclists are not flat, so lidar tells you it's not a cyclist.
> Meanwhile, Waymo is somehow going to map out in full detail entire countries in order to ever scale up.
They're owned by the preeminent digital maps provider, who likely has other uses for that data and will be able to package and re-sell it to the other vendors that require mapping data.
It's also worth pointing out that population, travel routes and income are not equally distributed. It's almost certainly an 80/20 problem where mapping 20% of a country lets you handle 80% of the rides. They can reach ~10% of the US population by mapping a number of cities you can count on your fingers. Highways are incredibly easy to map, so it wouldn't be difficult to interconnect those cities.
Again, it's just different approaches to solve the problem.
99% of interventions I had were from a driver's experience perspective, not the actual safety.
It's still far away from 99.999 that you would expect.
It seems silly to analyze a 4 year old version of something that is changing extremely rapidly.
Both Waymo and FSD have come a very long way since 2020.
FSD 12.3 is the current version of Tesla's self driving software. The article is comparing that current version of FSD to Waymo's 2020 state and saying that Tesla's self driving code today is worse than Waymo's 4 year ago. That is, according to the quote, Tesla is more than 4 years behind.
if you want to drive across the ultra straight highway flyover states it's game changing. if you don't do that, it's not that useful.
I have watched enough recent Tesla self-driving ride along videos on YouTube to suspect you might be mistaken on this point. Tesla intends to launch a cybertaxi fleet and their software looks like it will be good enough to get them there without lidar or additional sensors.
It is one thing to cherry-pick flawless drives on a sunny day and upload it to YouTube while having someone behind the wheel ready to take over the glorified driving assistant system. It is another to run a commercial driverless service open to the public 24/7 in one of the biggest urban areas, knowing that riders will record everything, assuming accident liability, and keeping a nice safety record without someone behind the wheel.
Tesla FSD sucks extra bad on sunny days in fact, due to its basic optical systems.
I just watched the latest video from AIDRIVR on YouTube. AIDRIVR is a TSLA pumper-and-dumper who has dedicated their channel to uncritical praise of FSD. In the first third of the video FSD v12 runs two stop signs, once directly into oncoming traffic in a 1-way traffic control and once at a stop where the cross traffic does not stop. This stuff is not even a little bit ready for fully supervised operation. https://youtu.be/fpoXr_z_6a4?t=565
Thanks for sharing. I skimmed through the video and watched a fair amount of it. I got a different impression.
I thought it was impressive how FSD 12 navigated narrow winding roads with parked cars and oncoming traffic and flaggers holding signs that alternate between stop and slow. My impression was while it's not perfect, it's a few iterations away from having very few situations that require disengagement. And keeping in mind that every incident of disengagement is a learning and improving moment for FSD, the following iterations of FSD will continue to get more impressive.
This isn’t just “a few iterations away”.
Most of what you said has been repeated for years from people who just watch curated YouTube videos. I’ve driven on FSD v12 and I’ve intervened multiple times almost every single drive. It’s nowhere near ready and will likely never be with that sensor suite.
I suspect that Tesla is paying attention to the disengagement events and working hard to minimize them in the future, but I truly have no idea.
Also I am sure this question has been asked before but what is good enough for FSD? Perfect in every situation? Better than the average human driver? At par or better than an expert professional driver? I don't have an answer personally but I am curious what others think.
What’s good enough for FSD is being able to do it without a driver present like Waymo does. Their crowdsourced reliability data in https://www.teslafsdtracker.com/ suggests they need at least 3 orders of magnitude improvement to remove the driver.
A liability-viable self-driving car needs to be reliable enough that you would expect to see zero significant errors in a typical journey. That's around the point where you will only have a few articles about one of your thousands of cars going wrong each month. Commercially viable needs better than that.
Don’t even get me started on the “didn’t take psych 102: Attention and Memory”-level cluelessness required to believe a human can safely pay attention well enough in a vehicle that reliably tricks you into believing it’s autonomous to take over in the split seconds before a disaster…
I find it hard to believe that the Tesla and Auto Manufacturer positions aren’t knowingly deceptive. I mean, what are they going to say? “It’s too hard so we’re just waiting for Waymo or Cruise to license their tech once it works”?
I’m gonna stop here before I start mocking geohot… I seriously can’t believe the journalists who wrote those early stories were willing to risk their lives like that…