The above is situations I know of offhand hand were my vision isn't up to the task yet I do anyway - because I don't really have other options. If other sensors can outdo me then I want them.
Not sure how to say it, but try to never do that going forward, please, for your sake and for the sake of the others who share the same road with you. Yes, driving "blind" for one second, maybe two seconds (even though it's uncomfortable for me at this point) from time to time is not the end of the world, particularly if you're driving straight and you had "scanned" the road in front of you beforehand, but driving while "trusting" in others will get you into accidents. Never drive faster than you can see.
So as an experiment, at 70 mph I put my face down on the gas tank: I could see the highway lines on either side, but literally nothing ahead. I wanted to see how long I could hold that position without freaking out. Answer: about twenty seconds.
This is far from the only experiment I did back then, I’m lucky to be alive.
For others left wondering what this meant, I think GP meant "I often drive with [the sun] in my eyes at dawn/dusk .."
This is still not 100% safe (especially in blind corners), but this is the minimum One should do.
And what? Be the lead car of the N-car pileup?
It's real easy to spit out idealized "never/always" type crap on the internet but meatspace is full of nuance and no rules of thumb can save you from having to make judgement calls.
I have driven on the motorway in heavy fog, and like everyone around me I slowed down, and there were no accidents.
Acting like an idiot because you think others will is not a good course of action, especially when you will cause issues to the majority who do not act like idiots.
I'm in the FSD beta and previously Tesla bounced the radar off the ground and was able to see a few car lengths ahead, vision only, not so much. The effect is when stuck in heavy traffic on surface streets FSD beta will think it should try to go around a lead car that is stopped, not knowing there is a long line of traffic ahead of the lead car. I don't see how they can overcome that issue with vision only, especially given current camera placement.
Could you do this with cameras? Possibly, but it would be a lot of cameras and a lot of mechanical parts—not ideal for a car.
Using more and better sensors is an attempt to compensate for the shallow mental model, so you'd have cars still being 'dumber' than human drivers, but if they're all-knowing, that doesn't matter quite as much.
This is a decent thread that has some visuals, some others have created 3d models demonstrating the blind spots the cameras have.
As far as Tesla goes, I disagree with their decision to remove sensors from their cars, however I also don't think Lidar, Radar, etc. would help all that much. Radar isn't high enough resolution for driving through dense fog, and Lidar doesn't work in fog. Driving in fog is just hard.
If anything, I think we've seen that computers are much better than humans at making decisions in an ever-expanding set of narrow contexts (eg.: chess, go, protein folding...). It's not so much a matter of "can a computer do it better", it's more a question of "when are we going to figure out how to break down the problem in a way that a computer can solve much better than a human".
But consider how a human or even a dog catches a ball. You can study physics and learn that a ball will follow a parabola when thrown and the the shape of parabola is determined by the velocity as it leaves the throwers hand. But none of that matters. You try and catch a ball. Miss a bunch of times. Practice. And then you get better at it. And then you can catch all manner of things coming at you at various velocities. You can unconsciously predict a ball's path accurately enough to catch it without even knowing any physics. Some people can even rely on their instincts.
This line of thought is moving goal posts, but I think even for people that have some understanding of what AI is, the current state of the art doesn't pass our intuition about all the things a human does to drive a car. And the sorts of mistakes that current self driving makes don't really seem better than a human.
The possibility is not in dispute. Billions of biological organisms do these things everyday.
But this is:
>It's pretty much how your brain solves these sorts of problems too, there's no magic to intuition or human decision-making.
Is it how the brain works?
My ball example isn't about it having a closed solution. If you throw a ball I can track it with my eyes and I know it's going to follow some path. If there is a gust of wind mid flight I can even attempt to make corrections on the fly. Someone with no notion of projectile motion can do this and never come up with the concept of projectile motion.
The "hardware" so to speak figures this out on it's own and the same "system" can be applied to infinitely many problems without prior knowledge and get results. This would be like showing a Tesla some chess games, and then the car realising, this is some sort of game, and then learning how to play or inventing something new to do with the pieces.
The results so far are really terrible compared to our intuition. We have Teslas that drive straight into road barriers and can't predict that a person might reappear after moving behind an obstruction.
This is even more goal post moving and I admit it's very unfair. But I'm not baby sitting a car.
(What is a mystery is the exact way things are wired up, but not the building blocks, and not the general organizational principles.)
Source: I did neuroanatomy in med school (but am not a doctor).
You don't think that's a lot of hand waving? Can you explain why the details wouldn't matter?
I don't want to mislead, there is an incredible amount of stuff we don't know... But we do understand more than most well-informed people think, even people in adjacent fields.
IMO, it's 16+ years of life experience that a computer doesn't get. Sure we train computers on what things likely are in pictures, and where they are on some sort of map. But that gives zero context into the 50,000 other things going on at any moment in a busy street.
I really wish transit was better. Even in cities that have what we call good transit, it is still bad enough to too many people drive. (Note, the above issues with human drivers apply to bus and tram drivers.)
there are plenty of cities with great transit, just not in the states
Even in cities you consider to have great transit there is room for improvement.
They are an engineering company, and they will attempt to engineer their way out of the wrong decision (vision only) if they have to in order to deliver.
[1] https://electrek.co/2021/01/13/tesla-millimeter-wave-radar-e...
[2] https://www.ti.com/lit/wp/spry328/spry328.pdf (Page 3-4, rich point cloud data, improved velocity resolution from 60GHz sensors, looks a lot like the benefits you'd get from LIDAR)
Companies do it all the time - I'm not even sure it's really lying (as they "collectively" believe they'll be able to do it) but they are willing to pivot when the time is right.
Apple's famously done it - the App Store was never to be at first.
You could imagine an AESA phased array at 60ghz. But there's a reason you only see those radars on aircraft or military vessels: you need hundreds or thousands of individual transmitters. That's why they cost millions of dollars and draw kilowatts of power, it's not (entirely) defense contractor featherbedding. So where's the cost savings over LIDAR? You can drop the number of radios to save money, but then you don't have meaningfully more resolution then any other automotive single-point radar.
Wouldn't you have to keep that on your books as liability in the same height as the income you had from the sale (as the true cost is not yet known)?
I thought that radar was inside cabin only, to detect kids, etc.?
Tesla is not operating in a way any ethical engineer should tolerate.
Let they who is without engineering sin cast the first stone. Intent and risk appetite is a spectrum.
[1] https://en.wikipedia.org/wiki/General_Motors_ignition_switch...
Death is unavoidable, only preventable short term on a scale. To expect zero deaths is unreasonable.
I won’t buy a GM product and I won’t buy a Tesla product either, happy?
Conflating the two makes it harder to have a conversation about what either actually is or does.
- Autopilot has been "on" during 12 fatal crashes. It's a level 2 driver assist so it isn't "responsible" for those deaths, the driver is. Autopilot data shows it improves safety. Could you explain the unethical part?
- The ignition recall, GM knew it was a problem, and it was. The unethical part is super clear.
These aren't equivalent at all. Actually this highlights the bad faith attacks on Tesla. There are reasonable controversial decisions to be discussed with Tesla's approach. But safety isn't one of them, they have the safest cars on the roads and the lowest crash/fatality rate.
Autodrive is going to force the actual law and the theoretical law to come into sync, otherwise automatic driving cars will cause other issues by sticking to the letter of the law when normal drivers do not.
For whatever reason, Tesla has decided to eschew this hard-earned wisdom, have treated their customers and the general public as an extension of their corporate research lab and their customers and others have paid the price for it with their lives.
During the DUC, contestants competed in an abandoned military base in the middle of the desert. In that environment, the vehicles were outfitted with all matter of flashing lights, emergency stop procedures, and all humans had to clear the area. That's what robot research was like in 2007. And our cars didn't have a verified track record of deaths!
Little did we know at the time, we could've just held the DARPA Urban challenge in the middle of downtown SF. Apparently we wouldn't even have needed to inform the unwitting public that an experiment involving a 2 ton machine that has killed people in the past was taking place in their midst.
Dan Amman’s original edict to Cruise was to Beat Uber. Uber was not just disruptive in 2016, but they really wanted everyone to feel the size of their ambition. E.g. they placed an order for 100,000 Mercedes S-Classes that ATG was going to then equip. Uber could raise billions easily, and they were culturally reckless, so the OEMs thought them a threat.
When Dan Amman took over, his pay package showed he very much wanted to be a Silicon Valley billionaire (see S-1 docs) who saved GM. And he had history with Mary. To Kyle, Mary is much closer to a boss, but to Dan, Mary is more of a colleague.
Fast forward to today and the self-driving landscape has changed a ton. Uber barfed all over themselves and COVID has changed the world. Dan Amman no longer had a case for making him a SV billionaire, and he never was able to fix Cruise’s culture problems* anyway.
So the change makes a ton of sense, and I’m sure Dan will find some other way to become as rich as sama or Justin Kan. Dan has already done a ton for GM and he’s good at what he does.
* I left Cruise a while ago and while Kyle hired a lot of good people, he also hired a ton of jerks. In some cases he did it personally. I have four different stalkers from Cruise, one who apparently slid into a panel and got me denied a job last year. Kyle, while I do think you’re the best CEO for Cruise right now, you created some culture problems that have caused lasting effects. Focus on what customers want and let the engineers get their job done.
0: https://twitter.com/WarrenJWells/status/1491487543455465472?...
0: https://youtu.be/ttvBZtqxlRo?t=630
1: https://twitter.com/greentheonly/status/1497696259473391617?..., and auxiliarily: https://twitter.com/greentheonly/status/1497696259473391617?...
Unstructured learning on Image+Text pairs has exploded and set STOTA on benchmark across the field. While the data coming from LIDAR may be better. There is no where near the amount of LIDAR data in the world. Images and cameras have the advantage here where internet is full of weakly labeled data that turns out to be the key for building model backbones that you can retrain for downstream tasks.
For example, if a a car is stopped at a red light and it sees a child across the street running and the child runs behind a truck it will take her 5 seconds to get past at her last observed rate of speed, and the light turns green at a moment when the car would hit the child as she emerges from behind the truck (given the preceding) as it drives by the truck, do any ML systems have the ability to predict this?
People would likely be cautious if they saw the same scene, and would approach slowly. I don't think that current ML systems are able to do that, but would like to hear some informed opinions.
About 2/3rds down the page features 'Object tracking' and 'Robotic control'
Let’s try a diverse set of strategies and see which one wins in the real world. Nobody knows for sure at this point.
Both of those reasons are of course flimsy, if true: form cannot always come before function.
I've always read it as Tesla trying to do a paradigm leap. It seems everyone in self-driving knows you can do it pretty well with a huge variety of sensors, but what if you got extremely good at only camera-based then added in the new sensors.
Basically, if you can establish that the very worst case works, you can add in more sensors to go beyond that and have the cars become superhuman rather than having to depend on all sensors to consistently be working to have a safe vehicle.
I’m not sure how valid an assumption that is, they could wind up just adding complexity and creating noise.
And what if the best answer is being extremely good at lidar-only and then adding in cameras on top.
now is a terrible time to buy a new Tesla car.
But I do realize not everybody can wait it outI think that's debatable. There are plenty of stories of trying to force an outcome of a big software project and things not going as expected.
https://medium.com/@herbcaudill/lessons-from-6-software-rewr...
Naked cameras jutting out are no better.
Re: dirty windshield. Humans see through those, and we have windshield wipers if visibility is impeded.
I'm still excited about the LIDAR possibilities but vision alone has done a remarkable job.
Not for Tesla. So many interventions. So many situations where the car actively tries to kill the driver and pedestrians.
It's strange to me that Tesla who are building this in the open are criticised for being dishonest. Meanwhile ALL the competitors are closed door private development and some how we should trust them more? In an industry that is known for deceitful behaviour.
Qualitatively, there's no comparison between them. Waymo and cruise are today more than capable of autonomy without a driver in certain circumstances. Tesla still has trouble staying in the lane and doesn't define an ODD or even claim L4 for fair comparison.
For background, I've been in waymo and cruise' vehicles, and ridden with AP. I've only seen FSD in video form.
Tesla publish autopilot data every year. Crashes per mile with AP, without AP, and with Safety assistance features, and show consistent improvement of AP. What do the competitors publish?
You are comparing AP (a traffic aware lane keeping cruise control system) to Waymo (a driverless taxi)? How about FSD Beta vs Waymo how has your experience been there?
By "in public" I mean I can go on YouTube and look up "FSD beta version 8", and "FSD version 10" and see hundreds of videos comparing both unplanned and repeated routes and see improvements and regressions. If I'm in the US I can even get the software myself and try it out (after hurdles)... how much more public could they do it?
If your metric is simply public access and videos, waymo and cruise both have public demos available. jjricks has quite a few videos on the former. The latter is too new and small a program to have much content yet.
> They also don't have professional testers like cruise, waymo, etc do validating releases.
Source? From my understanding they rollout releases to employees first (I assume a subset of those are professional testers). Then to beta testers. There have been a couple of releases that have been halted that lend credence to that.
> They also don't publish FSD data, and their AP numbers are essentially useless for determining actual safety
So is this because you can't see detailed data of their in-progress beta software? Could you show me the beta stats of Waymo/Cruise? Where are they published? Tesla aren't operating any robotaxis - so it's clearly not a finished piece of software is it?
On the AP front, why is the AP data useless? Which data do you need that is missing?
https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...
I would love to see disengagement reports for autopilot and FSD. I can't go more than 50 miles without having to takeover for autopilot. Like hitting the gas when it phantom brakes. You can't even use autopilot at night any more due phantom braking issues. I don't even use Navigate on autopilot because it literally tried to kill me on multiple occasions by trying to change lanes into a concrete barrier. Want to see funny. Use smart summon. I wish i could use smart summon and my phones video at the same time.
https://twitter.com/TaylorOgan/status/1458169984987942913
> Cruise drove 41,719 miles in autonomous mode per disengagement.
> Waymo drove 2.36 million miles autonomously with one accident that was the vehicle's fault.
> In the 4th quarter, we recorded one crash for every 4.31 million miles driven in
> which drivers were using Autopilot technology (Autosteer and active safety
> features). For drivers who were not using Autopilot technology (no Autosteer and
> active safety features), we recorded one crash for every 1.59 million miles driven.
That's 1 quarter and the numbers dwarf the combined total of all the competition. It's also a safety report they put out every quarter rather than an adhoc tweet.
How many crashes have occurred on FSD Beta? 0. [1]
I agree that LIDAR is valuable and hope all these companies are all able make safe cars with their tech stacks.
They work without synchronisation to the cars driving next to you.
Aren't AI models already better at image recognition than humans?
(I'm not suggesting that using LIDAR is not an improvement though).
On some benchmarks, AI models are better at very well defined tasks like image classification (“label this image from a set of 8 labels you’ve seen before”) or object detection (“draw a box around all instances of class X in this image, where X is a very narrowly defined class”) They’re not even close to being able to understand unscene examples and parse out their meaning in a larger context the way humans can. (“Recognize that this object in the road is a human riding some sort of bizarre unicycle he welded himself, then predict how he’s likely to move given the structure of his custom unicycle thing”)
The bottleneck in AVs isn’t “perception” in the sense of image classification and object detection, it’s deeper scene understanding and abstract reasoning.
LIDAR is useful for a ton of other reasons - ground truth depth, high visibility at night, great for localization - and can detect if something’s an obstacle or not without having seen it before (but false positives can be an issue).
Broadly speaking, people who focus on Camera vs LiDAR are missing the mark and don’t understand that the real difference between big AV players and consumer cars is HD maps and remote support.
Is there some cost prohibitive licensing on it that affects adoption the way e-ink patents have limited product creation?
Such incredible confidence in that statement that there is not even a hint or need to offer up any evidence or anything at all. HN is amazing. Dunning Kruger rules here.
To paraphrase XKCD [0] - if you want to cause accidents, there are other failure mechanisms that'd work on humans as well
This is directly the opposite of what Tesla has mandated their engineers to do, which is to be 100% reliant upon camera systems only.
Somebody at Cruise commented that their intention is to drive down the cost of LIDAR units through economies of scale and better technology.
My personal belief is that the data acquired from LIDAR representing a at-this-moment-in-time snapshot of a vehicle's surroundings is very valuable, and Cruise is probably going down the right path with this.
Relying entirely on cameras only requires the full intellect of a human who can make snap judgments about what's going on in a scene (eg: a pedestrian wearing a black or dark blue jacket who is walking a black or brown labrador retriever across an unmarked crosswalk in Seattle level mid winter rain, at night time, something I literally saw just two nights ago).