So Tesla is now going back and adding ‘HD Radar’ to the mix… after claiming it was a net negative and removing it barely a year earlier. No one saw that coming. <sarcasm>
For a company that partially exists because of the economies of scale and energy density gains in better battery tech, it always shocked me that Elon thought LiDAR was too expensive. In the time that Tesla has spent failing to achieve anything beyond level 2 autonomy, almost everyone else is about to release commercial vehicles at level 3… and they use LiDAR.
They’re going back because the supply chain has eased up and their internally developed hi res radar (“Phoenix”) is reaching production grade, so they no longer have to argue Vision is superior to keep shipping units post pandemic.
My prediction was wrong though, see below. I assumed NHTSA would’ve forced a recall mandating a move back to radar. They didn’t, and Vision was good enough to receive the highest rating in Euro NCAP’s protocol. You don’t need radar for Autopilot (we have Teslas both with and without radar), but if they say it’ll help with FSD, that would be a reasonable thesis to put forth (without arguing the merits of FSD).
> Overall, the Model Y received an Adult Occupant Protection rating of 97%, a Child Occupant Production rating of 89%, a Vulnerable Road User Protection rating of 82%, and a near-perfect Safety Assist rating of 98%. The Model Y’s 98% score in the Euro NCAP’s safety assist category is especially impressive as these are related to features that utilize the company’s radar-less Tesla Vision system.
> In fact, Model Y’s Safety Assist rating is the highest score among the vehicles that have been tested by the Euro NCAP to date. The Model Y’s stellar Adult Occupant Protection rating is also the highest thus far under the agency’s 2020-2022 rating criteria. Overall, the results of the Model Y’s safety tests prove one thing: vehicles equipped with Tesla Vision are just as safe — if not safer — than other vehicles out on the roads today.
> They’re going back because the supply chain has eased up and their internally developed hi res radar (“Phoenix”) is reaching production grade, so they no longer have to argue Vision is superior to keep shipping units post pandemic.
In other words, they lied that vision-only is superior to having a radar, put out a blog post on how Tesla Vision is going to solve FSD and then sold a few hundred thousand cars without radar for over a year.
The article actually quotes Elon saying back when radar was removed, that current radar decreased snr, but a high resolution radar would be better than pure vision.
So no lie. This fits perfectly with what he claimed.
He claimed just cameras were sufficient for self driving and sold cars based on that, even though they were working on a high resolution radar. Clearly, based on this article they are not sufficient. He also said phantom braking issues would be solved with radar removal. That didn't happen.
How many people would have bought the car (or FSD) if they knew a better radar was coming a year down the line that not only helps FSD but also regular Autopilot?
Wrong analogy. Not many would buy an iPhone if Apple removes the main camera and claims they can magically take pictures from the front camera alone, but reveals they will have a better main camera for next year’s model.
Elon has said it himself. There is never the right time to buy a Tesla. They are always changing things. And next year‘s car will always be better than this year‘s. And to some extend, this is true for anything you can buy in technology. So the idea that „if only people had known“ doesn‘t make any sense. The unspoken criticism here, of course, is that Tesla knew that they were going to re-integrate radar just a year later. And that they were withholding that knowledge from prospective buyers so that they wouldn‘t hold off making the purchase (there‘s a term for that effect, which unfortunately eludes me at this moment). But there is a much simpler explanation. They simply changed their mind. I think it took balls for Tesla to say that radar wasn‘t needed. After all, just a few years earlier they had written one of their big blog posts, which was titled „Seeing the World in Radar“. For them to say, „You know what, we realized we‘ve been on the wrong path. Fusing those sensors is actually making the decision process worse and not better“ - I find that a bold move. And it‘s not like it came out of the blue. Elon had said many times before that many phantom brake events were the result of radar and vision disagreeing with each other. So that issue had been a huge problem for them for a long time. The idea that these people at Tesla are sitting there and thinking about how they can fool the public seems just really crazy to me. From the start, their mission has been very clear: To get as many electric cars on the road as possible. And you just need to look out of the window in order to see that they‘ve kept their word. I‘m really quite puzzled by the amount of suspicion leveled at Tesla on HN. Have they missed their own deadlines on features? Of course they have. Elon promised the first coast-to-coast fully autonomous drive to be accomplished by the end of 2016. He literally said that he has „egg on his face“ in that regard. But it‘s them who said that goal in the first place. And it‘s clear that they are working on the problem, and that they are making progress. Elon Musk has managed to build a rocket company. His team made reusable rockets work - a thing that, to this day, no competitor and no government has been able to do. How can anyone seriously think that this man thinks about fooling the public? If you have power over the physical world - have seen that your team can make things that never existed before - why would you care about having power over people? Power over people is interesting only to those who have no power over nature. It‘s only when you cannot build that you become interested in stealing.
And about those „unrealistic“ deadlines… Tolstoy said something about that. He talked about Christianity. But it applies to big aspirations of any kind. I don‘t have the exact quote. But he said that if a man want to cross a river with a strong current, and he wants to reach a particular point at the other shore, then he cannot afford to just swim towards that point. Instead, he must swim towards a point that is far upstream of that point. He has to tell himself that he want to reach that point upstream, just in order to have a chance of arriving at his destination. It‘s not that he has to fool anyone. Or that he even had to fool himself. It‘s not that he has to swim „as if“ he wanted to reach that upstream target. He actually has to expend the effort it would take to reach that upstream target, and more. And I think that is what Tesla is doing. And Elon has said several times that it just isn‘t possible to communicate two different set of goals to those working inside the company and to the outside world. He has to focus on the target that seems just at the boundary of possibility. And when he does so, he cannot turn around to the consumer and tell them a different story full of disclaimers. He has to lead, and he has to burn his ships. Now, I understand that there are also Trevor Miltons in this world, and Elizabeth Holmes‘. But to put Elon in the same category just because both types of people demonstrate optimism in the future is, in my view at least, a pretty lousy heuristic.
> He claimed just cameras were sufficient for self driving
The latest builds of FSD beta prove that to be true.
The question now becomes what can be done to make it better, and how much better is required.
similarly, humans can clearly drive vehicles "sufficiently" after consuming alcohol. Countries disagree on the amount, and the definition of "sufficient".
> The latest builds of FSD beta prove that to be true.
Unless the latest builds of FSD beta work without requiring a driver, it doesn’t prove anything to be true. To be a self driving car, it has to actually work without anyone in the driver’s seat. That’s literally the definition.
It has to be capable of doing it thousands of times to be considered safe enough to remove the driver. Just one drive without disengagement doesn't count. FSD is not repeatable currently, so it's not close L4 self driving. For reference, Waymo and Cruise have a disengagement every 30,000 miles. FSD requires orders of magnitude improvement in reliability to match that.
No moving of goalposts. “Sufficient” means you need to demonstrate L4+ self driving with no driver. You’ve shown L2 driver assistance. By your definition, any car doing basic cruise control is self driving.
Euro NCAP doesn’t assess FSD safety, which was the reason they discontinued radar. If they’ve now decided they need radar, it means they lied in their blog post. They wouldn’t need it again if they thought it was safe without one.
Tying one hand behind their back in a fight proves only someone's ego is willing to take unnecessary risks than use every opportunity available. Elon is that type of a--hole who would gamble lives to be "right".
Unfortunately, your trolling attempt is irrelevant to the facts such as being unable to see stationary objects in the road when traveling over 80 mph. Would you prefer a less safe and less capable vehicle to "prove" how "smart" you think you are while killing more passengers and bystanders?
I suggest adding something of value in the future rather than try to look smart because you only appear to be a biased fool.
Doesn't look to me at all as if you're interested in a discussion of sensors. My impression is that you have an axe to grind.
Hence that quote. It points to the observation that when people get defensive about another person's ego, it usually means they see themselves somehow in a competition with that other person. Jung would have called it projection. Girard would have called it mimetic rivalry.
None of that is an excuse for your dismissive attitude and name-calling, which really don't belong on HN.
Humans are also able to drive with vision only. Same sentiment as Hotz from comma.ai has.
I also think that radar will be a no brainer and make it better over all (night, fog) but I'm not sure if it has to be a must.
Perhaps in 20years when you have overall very good car AI the non radar version is always behind but we are far away from this market saturation. Probably more than 20 car iterations and more.
> Humans are also able to drive with vision only. Same sentiment as Hotz from comma.ai has.
I’ve said this before and I’ll repeat it here. A CMOS sensor is nothing like the human eye both in terms of HW and the SW (ignoring the AI that interprets the signal - I’m talking just in terms of the sensor). Your eye is constantly subconsciously scanning an insanely wide FOV so that not only do you have an insanely high-res image at the center, you have a fairly high fidelity larger FOV that is quite wide. Additionally, the camera will automatically shift its focus insanely fast if there’s any motion in your periphery and start filling in the scene with extremely high fidelity data. It’s hard to really equate the two but some analyses I’ve found on Google estimate the human eye at having hundreds of megapixels worth of fidelity. The Tesla has ~8 1 megapixel cameras.
As for the intelligence behind the camera in terms of command and control, the human brain is significantly better.
At some point in the future, sure. A pure vision system could be equivalent (actually better) to a human driver. Today it’s probably better in some ways (doesn’t have the attention failure modes that the human brain can suffer) but fatally worse in others (distinguishing the white trailer on a semi right in front with a cloud or understanding there’s a dividor right in front of you). Obviously things are going to get better so I’d bet on a pure vision system in a decade or two will probably work. The problem is Musk’s false advertising and the lives it’s cost.
The probability of safety will be higher with pure vision than vision+radar, not lower. Vision has become so good that radar actually reduces signal/noise.
A very high resolution radar would be better than pure vision, but such a radar does not exist. I mean vision with high res radar would be better than pure vision.
I will be interested to hear when if/they announce this if they quantify how it is used, what exactly it improves, and will it be retrofitted on FSD vehicles.
The supply chain. They just needed to con people into believing getting less is good. The other automakers dealt with the supply shortage by delaying car deliveries... by a lot. Not an acceptable solution when you have an overpriced stock to pump even higher.
It’s a fun narrative but it’s bullshit right? The old radar and the new radar they are putting in are not the same part. So “supply chain” is the wrong answer.
The new radar has been under development at Tesla. The old radar didn’t have the resolution to be useful. Elon was actually clear from the start that they needed to build a better radar that didn’t currently exist for it to be useful.
But if you hate Elon it’s easier to call it a con than to understand the technical details.
> Since 2016, Tesla has claimed that all its vehicles produced going forward have “all the needed hardware” to become self-driving with future software updates.
The specific claim in 2016 was that "as of today, all Tesla vehicles produced in our factory – including Model 3 – will have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver":
I'm coining a new term: "gas-ware" New technology appearing in advanced electric vehicles. Much like 'gaslighting', a clever huckster keeps saying, its 'coming soon', and 'its has ___-capacity'. A common strategy for a 'gasware' creator, is he keeps moving the goal-posts. A necessary pre-requisite: high charisma and notoriety.
Thank you for mentioning gaslighting. I had that thought immediately but was wondering to myself whether it applied here (which means it was working!)
It's baffling that a company can do that and expect to still be trusted in the future, and with people's lives, no less. Alas, it'll likely blow over, if it even makes any kind of splash.
It'll be interesting to see the copious amounts of doublethink in the announcement.
To me FSD is a lie. I don't see why he would not use another lie to cover it up.
To safely drive on the road, we'll need to understand the environment surrounding us precisely. Unfortunately for all auto-driving approaches, I cannot see anything to do with that. Sensors are actually least to worry about.
Given the popularity of the Tesla, I would expect dramatic increase of (fatal) accidents involving Tesla/FSD in the next a couple of years. Hopefully I'm wrong.
How do self-driving cars do in the pouring rain or snow?
This "faith in technology" is borderline religion without proof, it's ignorance that it is as good as the code bugs and sensor fail (and people overriding safeguards out of laziness).
Pedestrians and cyclists are the "beta testers", I am sure the people in the hardened car frame will be mostly fine.
Yeah yeah, automation "will kill less than humans". Who goes to prison when software kills someone?
I can answer that because I just drove home in the rain at night in my Tesla Model 3.
It was a light rain that has been falling for hours. The roads are sleek but very few deep puddles. No problem with traction, but the sheen of the water combined with poor lane markings means it’s often hard to see the lanes.
My experience has always been the vision AI can discern lines and road markings significantly better than I can. So I often choose to drive with autopilot on and keep close attention at night in the rain.
The car will be almost always perfectly in lane while I am freed up to pay more attention to the greater surroundings, traffic signs and lights, route planning, and watching for hazards and other traffic.
I find at night in the dark driving without autopilot, lane keeping requires significant focus, which can mean other driving tasks like route planning and identifying hazards can suffer.
As of right now, exceedingly few cars are being driven entirely automated. Humans using autopilot will have more time and focus to identify and avoid cyclists/pedestrians that are on or entering the road. Autopilot is a massively useful safety tool.
Unfortunately some humans abuse their driving privilege. Drunk driving and distracted driving are huge and growing killers. Human drivers that are supposed to be paying attention that instead hit a cyclist / pedestrian are at fault. If it’s done while driving distracted on autopilot that’s even more so the human’s fault.
I find that Model 3 FSD hugs the center line even on roads with a huge amount of lane space, to the point that some cars and trucks will honk at me because they think I’m about to cross over. It also gets confused constantly by badly marked roads and can’t drive for more than a mile or two on complex streets without making some critical error or behaving like a drunk driver. It’s interesting to drive if you watch it like a drivers ed teacher teaching a vision-impaired student’s first lesson. It is not yet an assistive technology (outside of highways) that anyone should be delegating any amount of their attention to, and I wouldn’t want to be walking on a road with people using it that way.
I’ve been in FSD mode with this exact thing of being too close to the line, it’s dangerous. I wasn’t in the driver seat btw. Anyway, one time we were in a low density town and the Tesla swerved from being close to the center line to literally turning into an oncoming vehicle. Out of nowhere on a sunny day. Luckily my driver regained control but that scared me out of wanting to use fsd for years. Scared the shit out of the other driver I’m sure.
I don’t have this issue at all. AutoPilot (not FSD Beta) is not designed for back roads or complex city driving. It’s best on highways and state roads with limited intersections. I do not use it in the city.
The issue I have with AutoPilot is it doesn’t always shy away from poorly driven cars and large trucks in the next lane over. So while a human would be all the way over to the far side of the lane, AutoPilot stays perfectly centered or only slightly shys away from semi-trucks.
The current liability framework isn't designed to handle such a question.
A self-driving car surfaces the problem in a visceral manner, but this applies to many many sectors and to any software that runs critical services/tools/infrastructure - Hospitals, Medical devices, thermostats, etc.
Also, this covers physical harm, but what about economic harm? If an OS security bug was exploited by a hacker and caused you to loose all your life savings and end up homeless, can you put a linux/windows/ s/w dev in jail?
As the article points out, it remains to be seen if this is for self-driving use or in-cabin. Europe requires interior intrusion sensors in passenger vehicles, which afaik can be radar-based. There are also possibilities for fine-tuning of airbag deployment based on occupant size and shape which could be determined by a radar.
Sets off the alarm (an in a Tesla, gives you a phone notification) whenever a window is broken, or something enters the vehicle while it is locked. In an American car (at least last time I checked a few years ago), the alarm only goes off if you use an interior handle while the car is locked.
This is one of the most obviously false arguments in this domain and yet it is repeated often.
You don’t only have two eyes, you have a general intelligence which is far superior to anything that Tesla or anyone else has developed as a control system. The deficiency of the control system compared to a human brain means the sensors need to be significantly superior to anything that humans have in order to build an equivalent map of the world.
Even just taking the argument on its surface, electronic cameras are not a match for human vision:
* our eyes have incredible dynamic range. We can perceive details in deep shadow and bright sunlight at the same time in the same scene. Cameras aren't good at that, and certainly not at a price point a company that forgoes lidar for pricing reasons would be happy with.
* our eyes combine wide angle low resolution perception with high resolution capture of a single point of interest, and switching between points of interest is super fast. No-one's even trying to replicate this electronically; the approach with digital cameras is to go for high resolution across the board, but this is nowhere near as high as what human eyes can manage in the centre of vision, and also because it results in a ton of data, SLAM applications tend to go for lower resolutions and more frequent images instead to help manage transfer and processing latency; however, this makes correctly parsing the scene a much harder problem than it would be either for the human or an electronic vision system of comparable fidelity.
* our eyes include a bunch of processing to help with feature detection and cross correlation (for depth perception) before the signal ever makes it to our brain, which helps us deal with the latency/bandwidth issues described above. This is certainly something we could attempt to replicate electronically, but, well, Tesla didn't want to pay for lidar either, so not seeing that happening tbh.
* our depth perception doesn't merely rely on cross correlating stereo vision to produce a depth field like electronic applications do. It synthesizes this with information such as the angle between our eyeballs and the shape the lens must take to bring the object of interest into focus, both of which vary depending on how far away the object of interest is. (This is part of why VR helmets make many people feel ill: they supply stereo images in perfect focus a small distance away from the eyes, and so these sources of information all disagree with each other; therefore, motion sickness.) Without some equivalent source of information - some form of electronic rangefinder, like lidar perhaps - the AI, unlike the human, is forced to rely only on cross correlating stereo images, which is problematic in a variety of tedious ways.
I know - was just repeating the nonsense from Elon and his supporters club.
Our retina's does some major preprocessing and chunking it down for the visual cortex to handle in "realtime", which also feeds it into all of other parts to complete the whole picture of what we are seeing.
It took my daughter 2 weeks of tutoring and practice to master stick shift driving on the public roadways - FSD still working on it and sometimes it makes rather boneheaded decisions.
For a company that partially exists because of the economies of scale and energy density gains in better battery tech, it always shocked me that Elon thought LiDAR was too expensive. In the time that Tesla has spent failing to achieve anything beyond level 2 autonomy, almost everyone else is about to release commercial vehicles at level 3… and they use LiDAR.