From the last time that Cruise showed up on HN a while back, I recall I made a comment based on the photos of their test vehicles that they're clearly going all-in on LIDAR + RADAR + camera for feeding data into their autonomous onboard navigation computer.
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).
My problem with cameras is as a human I'm well aware there are things that I can't see and so i rely on faith. I've driven in dense fog trusting that nobody else would be stupid enough to do the same (fortunately taillights can be seen through fog, but not many other dangers that normally are not on the road, but if so they are dead). I often drive with sub in my eyes at dawn/dusk and trust that the road is clear. I come of the top of hills and trust there is nothing on the other side where I can't see until I'm over the top and it would be too late. Backup cameras have been very useful, and take away something I would have said a few years ago.
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.
> I often drive with sub in my eyes at dawn/dusk and trust that the road is clear
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.
A younger, stupider me was riding a motorcycle on highway 99 near Merced. It was just after dawn on a Saturday, with literally no one else on the road for miles.
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.
This gave me a sinking feeling in the pit of my stomach. I thought I was a badass letting go of my bmx bike handlebars for 5 seconds going downhill as kid in middle school… you really took it to the next level my friend. Glad you are alive and well!
It was very difficult to do. I still practice this, only while walking down hallways or alleys where the worst possible outcome is stubbing a toe (or maybe banging my head on something). Interestingly it doesn't seem to matter much: 20 seconds is seemingly my limit walking or motorcycling (although I don't know what my walking limit would have been back then).
>not sure how to say it, but try to never do that going forward,
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.
Or just don't drive. If you are driving blind like the parent is claiming they do (driving without seeing before you can stop is driving blind) then it is only a matter of time until you hurt yourself, or worse someone else.
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.
Humans can turn their heads. We also have stereoscopic vision that is able to gain visual clarity with minor movements. Our eyes also mechanically adjust the amount of light they receive, can focus dynamically, and are self cleaning.
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.
It's probably not, but humans have a mental model of the world that cars don't, even ones with relatively advanced computers.
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.
i hope you at least slow down around blind bends and corners. that shit is scary. see people racing over such obstacles all the time. i assume there will instead be a crash or deer or corgi or child hiding behind every obstacle
The solution to most of your issues is to drive slower. You should be driving at a speed such that you could stop if anything unexpected is beyond your visual line of sight. If that means you can only drive at 10mph in dense fog, so be it. I personally refuse to drive on a motorway in dense fog becuase there are always pile ups. Honestly, there are just conditions where we shouldn't be driving and society should accept that. Same reason we sometimes cancel flights, busses and trains.
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.
I don't think that was the parent's point. These things could be made safer with other detection technologies beyond plain cameras. A human driver can make a conscious, nuanced decision to make these "faith-based decisions", but a computer may not be able to. So perhaps it might just completely refuse to drive in those situations if all it has is a camera, but it can't see anything due to fog or terrain, even if it would be somewhat safe for a human to drive in that situation.
It's really not clear to me what gives a human the nearly-magical ability to make decisions that a computer could not possibly make.
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".
Your examples are things that can mostly be broken down or modeled logically though. Chess are Go are hard so I guess it seems like AI for them means there some sort of fundamental break through.
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.
I absolutely think the type of decision we're talking about can be modelled. I'm not suggesting it's trivial, and I'm not saying there's a nice and tidy closed analytical solution, which is what you appear to be focused on with your parabola/physics-based example, but I absolutely think statistical methods (or "AI/ML" if you prefer) can be applied to the problem. It's pretty much how your brain solves these sorts of problems too, there's no magic to intuition or human decision-making.
>I absolutely think the type of decision we're talking about can be modelled.
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.
Yeah, I concede that it absolutely is a lot of hand-waving. But we're also talking about emergent properties: it's possible to understand why they arise, what the emergent properties are, and how the underlying system works, without having full information about it all. In fact that's a very analogous situation to our thermodynamic understanding of gases, or fluid dynamics.
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.
I think there's some niche driving scenarios that humans will continue to beat computers at. What comes to mind most readily are meta-knowledge things like "It's 2:45 and I saw the school bus go by, right now it's letting kids off and blocking traffic up ahead so there will be no traffic on this cross-street". Or "The football game on the radio is ending soon, I better avoid the streets near the stadium because there will be a lot of pedestrians".
> It's really not clear to me what gives a human the nearly-magical ability to make decisions that a computer could not possibly make.
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.
The problem with lidar + radar is that it can't distinguish between soft drive-able things (smoke, steam, paper bag, grass, water splashing) from hard collide-able things.
If there is anything about what I wrote that is unique to me it is that I'm honest enough to admit to them. Everything I wrote applies to everyone who drives more than a few km in their lifetime. They might not happen everyday, but they happen to everyone on a somewhat regular basis.
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.)
People slow down in dense fog, but they slow down to see taillights distance which isn't enough for anything unlighted. People don't slow down for sun glare - that would be dangerous as the car behind you isn't - instead they assume there is nothing there, or they can figure out something is happening from the other lanes that aren't in the sun glare.
Even in cities you consider to have great transit there is room for improvement.
I would not take Elon and Tesla at face value. While they are currently going all in on cameras, I wouldn't put it past them to 180 and decide "turns out we need new hardware, and we're going to eat the cost considering our revenue and market cap." They recently filed to test new millimeter wave radar equipment operating at 60GHz [1], so I wouldn't say they're married to vision only if it turns out vision only isn't going to work.
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.
[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)
Yeah it has all the hallmarks of “we can build a fully automated plant - well turns out that humans are still way better at some things”. I hope they figure out if this path is a dead end, at least in the short term, or not sooner than later though.
This is patently what will happen when the time comes (if needed) that I'm surprised more people don't realize it.
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.
That electrek article doesn't make it very clear, but Ka band radar is not new. Project Soli used it for gesture detection. Those roadside "speedometer" signs run at 20ghz. Etc.
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.
Yea, it’s possible they are going all-in on vision only right now because LIDAR is just too expensive right now to roll out to their fleet. If they get it to work without LIDAR, then that’s great. But it’s possible they’re building out a system that can quickly pivot to LIDAR if necessary and when LIDAR costs come down to make it economically feasible as well.
More importantly if the main part they’re working on is the “what to do with situation X” code then adding additional sensors for more accurate data later is relatively easy.
The cost of retrofits will eventually be covered by the $6-12k everyone has paid for "FSD", if they can drag this out until solid-state LIDAR is cheap enough. And meanwhile they get to book that revenue as pure profit.
The milestones they've been using are bullshit - trivial and/or half-baked feature releases (e.g. stop sign and traffic light recognition) that are not reflective of the real challenges in shipping vision-only L5 autonomy. They already recognize half upfront, which is nuts.
To add, they also often have Lidar vehicles roaming around their Fremont factory and design center in Hawthorne, so they're definitely not ruling out any specific technology if it happens to provide measurable improvements at some time.
I thought the same, but have seen multiple product sheets and marketing materials from various manufacturers of this equipment describing it being used for exterior scene building and object discrimination (with about 40 meters of range). Continental (well known for its forward looking radar for automatic emergency braking) has a similar product that operates at ~77GHz.
There's a word for engineers that program cars to break the law resulting in deaths, and one of the first law codes specifies the appropriate punishment.
Tesla is not operating in a way any ethical engineer should tolerate.
GM, who owns Cruise, paid out claims for 124 deaths due to their ignition recall that they were aware of, did not disclose to regulators at the time, and had to forfeit $900 million to the United States government in 2014 [1]. Tesla's Autopilot has been attributed to 12 deaths [2].
Let they who is without engineering sin cast the first stone. Intent and risk appetite is a spectrum.
When quoting those 12 deaths, don't forget that
1) Autopilot is a driver assist feature (like cruise control, driver is always in control and at fault)
2) The majority of those incidents had a driver under the influence
I don’t believe Autopilot, a robust driver assist system with substantial safeguards, is irresponsible (as someone who has used it for ~60k miles of travel). I do believe allowing people to die because you don’t want to replace a defective ignition switch is. That’s not whataboutism, it’s pointing out hypocrisy. NHTSA believes Autopilot is safe enough to continue to allow its use on public roads. What makes a random HN participant a superior authority on the topic than them? Is it materially unsafe because someone here asserts so without backing up their claim? These are not rhetorical questions. 3700 people die per day from auto accidents that are ~95% human error caused and 12 deaths total while Autopilot was active is irresponsible?
Death is unavoidable, only preventable short term on a scale. To expect zero deaths is unreasonable.
The big problem with Autopilot is the sleazy way it’s been marketed and sold to customers and regulators versus what the system really is. Charging folks $10k for an option with the CEO insisting that “full self driving” is shipping next year is borderline fraud.
I won’t buy a GM product and I won’t buy a Tesla product either, happy?
- 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.
I think they’re referring to the “rolling stop” code Tesla added - which is technically an illegal maneuver but not one resulting in death as far as I know.
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.
When programming an AV to drive like a human does, are the programmers including all the contextual human judgement that goes along with the human decisions, or just throwing away the letter of the law because it's too much trouble? I have a hard time explaining to a human teenager the etiquette that goes along with something like merging onto a highway, and it varies by region.
That's because Kyle competed with MIT in the DARPA Grand Challenge, which is where the architecture for the modern driverless car was battle tested. LIDAR was by far and away the enabling technology in those competitions. Every team that finished in the DUC had a Velodyne 64 LIDAR. They are that important.
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.
This story is getting front-paged less about the tech or business aspects and more about so many in YC feeling validation through this outcome. But Kyle’s announcement is pretty minimal because it’s really more of a handoff than a validation of the YC country club.
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.
You may well be describing me, as a Seattle resident with a blue coat and a chocolate lab who currently has an upset stomach and is in need of many walks recently. We're working on the food/diet part, but your comment inspired me to buy a reflective arm band for myself. His harness is already a bit reflective. Anyway, thought it was funny and thanks for the sanity check/advice.
The thing is that I can't find any examples of Tesla's vision only system in FSD Beta not detecting any objects - all the failures are on the 'driving AI' side, ie. it makes the bad decision to try to turn right into a cyclist despite the cyclist being displayed on the center screen[0].
There are tons of examples of objects flickering in and out of existence, not being detected, warping in weird ways. This is one of the most recent tweets from a prominent Tesla reverse engineer with plenty of examples.
That's still detection, even if it can't precisely map the edge of the semi while cars are moving by; sometimes what's on screen isn't representative of what actual bounds the car thinks is off-limits, as we can see here[0] where the model of the commercial truck clips the car's model on-screen. He seems to support this theory that the problem is mostly the brain of the car[1], at least regarding how lidar won't help with object detection in general.
I used to agree with you, but I think the last 3 months in computer vision changes everything.
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.
I don’t think image+text pair models (E.G. CLIP) would be very useful for AV tasks. They’re not very good at fine-grained classification or counting the number of instances in a scene. Not even getting into latency or model size.
8+ responding distance dots that are much closer than they should be are already a STOP, APPLY BREAKS situation where not much ML is even needed. That's the beauty of LIDAR. It will save lives unlike Tesla's driving into obvious poles/truck booms mistaken for a statue of liberty in Nebraska.
To what degree can current ML models and techniques for self-driving understand what's happening based on data captured at multiple time points, maintaining object permanence?
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.
So is this a self-fulfilling prophecy? Some guy decides a trillion dollar company should discard an entire dataset for no reason, and as a result a decade late the company can finally actually do the job, thus proving the guy ?right?
I think Tesla actually used radar data to provide ground truth for this. So they don't even do that to generate ground truth. What happens if things have the same relative speed when doing the labeling? Or low light conditions? How do you account for per part variation in lens and sensor designs and how it messes with your predictions?
I do wonder if, deep down, Tesla's decision is rooted in aesthetics. The cars with all the crazy sensing equipment on them are ugly as sin. That really doesn't jive with Tesla's brand at all. A more charitable interpretation may be aerodynamics; making those worse would reduce the car's range.
Both of those reasons are of course flimsy, if true: form cannot always come before function.
> I do wonder if, deep down, Tesla's decision is rooted in aesthetics.
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.
But it's not finished. Have you asked Cruise testers how many interventions they have had to do in equivalent drives as Tesla FSD Beta?
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.
I'm not sure what you consider open about Tesla's approach. The fact that you can run updates without a license or any clue about what's changed or how it will perform in your area is exactly the issue! Tesla goes out of their way to avoid publishing actual metrics for safety, even useless ones like the annual DMV reports. Waymo, cruise etc aren't much better in public, but maybe it's a small comfort to know there are advocates internally (myself included) at all of these to publish that data?
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?
I would consider a reasonable definition of 'open' to include publication of all relevant safety metrics through the appropriate channels. As the other reply mentions, one channel would be the DMV report in CA which Tesla famously eschews. They also don't publish FSD data, and their AP numbers are essentially useless for determining actual safety. They also don't have professional testers like cruise, waymo, etc do validating releases. Those other companies can do better here with publication, but they track safety data very closely internally, even down to specific scenarios and streets/areas. I haven't seen evidence that those sorts of numbers exist at Tesla or are consulted before releasing updates to the public, but feel free to correct me.
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.
I'm sorry I don't consider a "curated CSV for California only" to be at all equivalent to 60,000 members across the US of the public who are free to post videos and setup failure cases, and report whatever they like.
> 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?
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.
> 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]
Cars are already complex enough to make them vulnerable to global supply chain challenges. The more tech that's required to put a car on the road, the more challenged that manufacturer is going to be have enough inventory of all the components to actually ship cars.
I agree that LIDAR is valuable and hope all these companies are all able make safe cars with their tech stacks.
Don't you have a big problem with interference? I wouldn't trust a Lidar/radar based vision if there were 30 other cars waiting at an intersection. Yes, you can use multiple bands etc, but that's the advantage of using more passive sensors such as cameras.
They work without synchronisation to the cars driving next to you.
> Aren't AI models already better at image recognition than humans?
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.
The sort of abstract reasoning I’m talking about is beyond the capabilities of any ML model that will run onboard a car - the “abstract reasoning” problem in AVs right now is solved primarily through HD mapping and remote assistance.
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.
LIDAR data contains highly accurate depth information, at which point you don't need the abstract reasoning nearly so much. You can at least do basic object tracking and collision prevention without it.
I've been out of it for a while, but my understanding is that humans are still better at video recognition and extracting high-level information from moving videos, while the state of the art CV still tends to focus on rapid frame-by-frame static classification, with some semblance of persistence and motion strung together at the end.
There are certain tasks where the models can do better than certain kinds of human effort (e.g. can do better than a person who is tired from doing a thousand of these in a row, is paying very little attention, is spending very little time, and doesn't really care), but that hasn't translated into actually doing image recognition better than humans.
LOL no, not even remotely close. ML image classifiers will decide that an otter is a bicycle with some invisible (to humans) noise injected into the chroma. And that's state-of-the-art classifiers, which Tesla does not possess.
> 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.
You've misunderstood his sentence then if you think the question mark is justifying the e-ink claim about patents. And if you look at my comment history, you'll see that this repeated claim/statement about patents vs electrophoretic products is unsubstantiated and no one has been able to provide any factual support for it. I stand by my claim that HN seems to have a lot of Dunningites when it comes to electrophoretics.
I'm in Michigan and I've watched the auto industry close up for a long time. I've got friends at the Big 3. GM decided it was better to put an adult in charge of Cruise. Someone they trusted and they thought had a clue on tech.
Turns out it didn't work well. GM learned the hard way the founder was the best CEO of Cruise. There's a lot they do not understand about how startups and Silicon Valley work.
Kyle Vogt was removed from the CEO position in late 2018 after Cruise failed to make progress toward a number of milestones. The milestones were outlined as part of the Cruise acquisition.
Dan Ammann took over and made really substantial progress. Ammann is widely considered to be the reason Cruise is where it is today. His record of success was in sharp contrast to Vogt's consistent record of failure.
Unfortunately, Ammann disagreed with GM's CEO about Cruise's future as a business. Employees at my level of seniority don't know the details, but many of us are skeptical that Kyle is the right man for the job.
At least according to Tech Crunch it was because Ammann also failed to deliver something on Barra's agenda [1]:
And while Ammann continued to push the company to expand, there were missed targets, notably the plan to launch a commercial robotaxi business in 2019. The company has spent the past two years inching toward that commercialization goal, along with the rest of the industry, which has gone through a spate of consolidation.
Most likely that's a lot harder to get done than anyone thinks. Is Vogt really the guy to make it happen? I guess we'll find out. But GM sending an otherwise successful VP packing after more than a decade there signals they are looking to shake things up. GM is not Silicon Valley and even senior execs IMHO enjoy longer tenures that you'd find in the Valley. They also skipped over the whole reassigning him to another role / oh he's left to pursue other interests tap dancing you often see to give people some runway or cover to leave. Not sure whether to read anything into that or if that's just how Barra operates.
The failure to launch in 2019 was shortly after Ammann took over (late 2018).
Ammann inherited Vogt's mess, and at that time nobody in management fully understood how bad the situation was. Ammann promptly began fixing things and we see the fruits of that today.
Ammann's departure, to the best of our knowledge, is a result of a conflict with GM's CEO regarding the future of Cruise as a business.
Hopefully Vogt can preserve, and successfully build on, what Ammann accomplished.
fully appreciate if you can't go into detail, but what changed? It seems like they'd always been working towards robotaxi-ing from a product perspective so it must have come down to execution-level stuff, no?
what is Cruise's hiring process like atm? I'm very interested in the AV space as a new grad and have worked on a relevant project with a real vehicle, but never heard back.
I have no degree and got to an onsite with Cruise before cancelling for another role
Ironically I ended up canceling in large part because of the CEO spat.. it gave me some uncertainty about if they were committed to L5. Especially vs serving as GM's L3 supplier while paying L5 lip service indefinitely (and never going public)
If anything, I got the impression Dan had a pretty good handle on the technical aspects of Cruise and the path forward for the company, and that the GM CEO fired him because she didn't like what she was hearing from him. Dan wanted to pursue expanding Cruise robotaxis, while Mary Barra wanted to find ways to integrate Cruise's technology into consumer GM cars - that's not how Cruise/Waymo/Zoox-style robotaxis work, and any such integration would necessarily involve rewriting a huge portion of the stack.
There was also reporting that Kyle was offered this role immediately after Dan was fired but turned it down, opting to remain interim CEO while they searched for a replacement. Looks like something has changed, or they couldn't find a suitable replacement.
> [integrating] Cruise's technology into consumer GM cars [is] not how Cruise/Waymo/Zoox-style robotaxis work, and any such integration would necessarily involve rewriting a huge portion of the stack.
That sounds like a [citation needed] to me. Surely there are UI problems to be solved in a car with a human being in the driver's seat, but the sensors and automation decisionmaking is going to be identical, and that's where the hard parts are.
Or alternatively: a less charitable way of making the point work would be to say that Cruise's stack was a bunch of special-case hacks for specific regions and vehicles and wasn't scalable to arbitrary environments nor regular passenger cars.
I work in the AV space, so I guess you’ll just have to trust me. I assume Cruise internals look very similar to the internals at other AV companies, which I think is a safe assumption.
>The sensors and automation decisionmaking is going to be identical, and that's where the hard parts are.
Sensors will be very different. LIDAR prices are coming down but we’re still a while off from incorporating it into consumer vehicles - and we’re way, way off from incorporating multiple high resolution, long range LIDAR pucks, plus short range LIDAR to cover blind spots, plus imaging radar, plus thermal cameras, etc. into commercial cars the way they’re integrated into Cruise cars right now.
Automotive decision making will also be extremely different. Modern robotaxis rely on very high detail HD maps that are continuously updated. It’s impossible to scale that nationwide in a way that would work in a consumer car. Fundamental parts of the stack assume these maps are present and accurate - remove that assumption and a ton of things have to be rethought from first principles.
Remote support is also a key part of AV stacks - that is, asking a human to clarify an ambiguous situation. There’s a video where Kyle says that Cruise cars request remote support approximately every 5 minutes. Again, fundamental parts of the stack rely on the availability of remote support and have to be rethought from first principles if it’s removed.
This isn’t even getting into the differences in compute on a Crusie car and a consumer car.
>Or alternatively: a less charitable way of making the point work would be to say that Cruise's stack was a bunch of special-case hacks for specific regions and vehicles and wasn't scalable to arbitrary environments nor regular passenger cars.
I mean you’re not wrong, but you can’t create a robotaxi without this. ML models just aren’t capable enough to handle edge cases in a safe way without tons of hacks in place.
> Automotive decision making will also be extremely different. Modern robotaxis rely on very high detail HD maps that are continuously updated. It’s impossible to scale that nationwide in a way that would work in a consumer car.
What are the challenges here? Is it simply the different scale required on the backend to handle tens (hundreds?) of millions of privately-owned cars, vs. that required for orders of magnitude fewer robotaxis? If so, I don't think that's all that insurmountable. I guess one thing that would worry me there would be bandwidth on that scale, as well as what happens when a privately-owned car is taken into an area where it doesn't have connectivity. I assume a robotaxi would just refuse to go where it can't talk to the internet, while that would be unacceptable for a private car to do. (Then again, the private car could still have manual drive controls, and require a driver to take over when internet connectivity is lost.)
I also work in the industry. You've got a couple misconceptions here about how AVs actually work. First, virtually all processing to determine what to do on the road is done locally, in real-time. Relying on internet connectivity would be a safety nightmare. This means you have to pre-train your models to deal with everything in a geofenced area before the car starts driving there. Hence the need for very high detail maps of the areas where robotaxis are going to run. Second, the concept of L3/L4/L5 driving doesn't quite match the reality of how things are playing out. Due to the limitations and cost of deploying current-gen self-driving systems, AV companies are designing vehicles to be L4/L5 within a specified design domain. That means a specific set of roads under specific weather conditions, light conditions, etc. Then those design domains will slowly expand until one day, years from now, vehicles are actually what the public now thinks of as "L4/L5" and at that time it will be possible to deploy the tech for privately owned vehicles. Until then, you're looking at commercial fleet vehicles in specific use cases only (robotaxis, automated heavy duty trucks, farming equipment, etc.)
But that’s not what Tesla is doing! You’re absolutely right and it’s also a safer route (for others on the road) to go with robotaxis first. This is precisely why you can’t focus too much on competitors. GM and their CEO (who seems to be really smart) is focusing on Tesla here and trying to compete with them.
Tesla isn’t focused on robotaxis - they’re focused on trying to sell a premium ADAS feature their CEO over promised half a decade ago. Whcih is necessarily designed very differently from a robotaxi. It doesn’t make sense to try to force a stack built for robotaxis into a consumer ADAS car because you’d have to fundamentally rewrite the stack anyways. This is why GM had its own ADAS team while Cruise focused on robotaxis. I’m sure they’re looking at Teslas market cap and hoping they can generate hype by trying to say they’ll integrate Cruise tech, but that’s just not how these stacks works.
That's exactly what OP is saying - GM is focusing on 'compete with Tesla' rather than pushing for progress/allowing Cruise to progress the technology on their own schedule.
Not going to give my name for obvious reasons, but formerly at Cruise. Any sufficiently complex system has lots of hidden assumptions that aren't trivial to unwind when changing problem domains. I can think of at least a few design decisions I personally made that will have to revisited for personal vehicles. It's not impossible, but still nontrivial.
GM was also ready to get into bed with Nikola which was an obvious pile of bullshit even before the Hindenburg report. I’d not want to be the one holding the bag there while management scrambles to catch up. Probably the best for him to get out now.
> Turns out it didn't work well. GM learned the hard way the founder was the best CEO of Cruise. There's a lot they do not understand about how startups and Silicon Valley work.
I'd be hesitant to generalize this lesson. Another case of "adult supervision", in that case against the wishes of the founders, was.... Eric Schmidt at Google. Say what you want about Google, but it's difficult to say that they weren't successful under his tenure (2001-2011)
I have a great deal of skepticism concerning fully-autonomous cars, TBH. There are just too many corner cases around weather and sensing, etc., not to mention insurance and other regulatory hurdles.
Personally, the idea of "a wire in the road" by which I mean something like cats-eyes in the road but with RFID tags placed every 40 feet, or some other common means maintained by the DOT that all cars can use to sense where they are, is neglected because every player in autonomous cars wants the "winner take all" method of being first, which will give them at least, a $25 billion market cap.
> Personally, the idea of "a wire in the road" by which I mean something like cats-eyes in the road but with RFID tags placed every 40 feet, or some other common means maintained by the DOT that all cars can use to sense where they are, is neglected because every player in autonomous cars wants the "winner take all" method of being first, which will give them at least, a $25 billion market cap.
You can safely assume most of the major players have thought through ideas like this and abandoned them for good reasons. In this case, you wouldn’t be able to rely on these external sensors being online with 100% uptime, so you’d necessarily be forced to build the capabilities to operate without them anyways. And adding external sensors and local communications and trying to deal with different standards and models and interop adds a ton of complexity.
I don't think anyone has seriously explored collaborating with DoT. If they had they would have at least talked about it.
> In this case, you wouldn’t be able to rely on these external sensors being online with 100% uptime, so you’d necessarily be forced to build the capabilities to operate without them anyways. And adding external sensors and local communications and trying to deal with different standards and models and interop adds a ton of complexity.
Well, yeah nobody said it was easy. But, it does seem like something that could work, whereas we know that fully autonomous is just not possible with the tech we have today.
Define “fully autonomous” - we have driverless cars in Phoenix and San Francisco right now - granted, with some pretty narrow constraints. Not as narrow as “needs custom hardware installed along every road to work”
Another thing to understand is that perception in the way you’re imagining these cameras would work isn’t really an issue for AVs right now. I don’t think the problem you’re describing is a real blocker for AVs.
> I don't think anyone has seriously explored collaborating with DoT. If they had they would have at least talked about it.
Collabing with DoT is a terrible idea imo.
Anyone who has seen anything any DoT in America has tried to do in the last few decades should immediately know why. It will cost an astronomical amount of money along with an equal time length.
To put "wires in road", it would likely be a project on the scale of trillions and 50+ year deadline. It won't happen, it's more feasible to solve the problem of how to have smart cars on dumb roads rather than relying on bloated and corrupt DoT & contractors to do the work of making smart roads.
This isn't even to mention the likely quagmire of what happened with early railroads where the rails laid by different companies weren't compatible with each other and you had to switch to an entirely different train to continue your journey.
> You can safely assume most of the major players have thought through ideas like this and abandoned them for good reasons
I'd LOVE to hear the reasons why this system won't work. If it is something other than all the big car companies refuse play nice with each other I will be SHOCKED.
Let’s assume your proposal — wires in the road to aid lane-keeping — works perfectly. What does it get us aside from lane-keeping? That’s far from the hardest problem. It doesn’t account for stopped vehicles, cyclists, pedestrians, deer, or any of the other things human drivers commonly encounter and avoid without fanfare. Even if we imagine some sort of car-utopia where only AV vehicles are allowed on your perfect roads it doesn’t account for common vehicle malfunctions like blown tires.
You shouldn’t IMO scope the problem only to lane-keeping. That’s a very small part of the problem space. Solutions aimed at only that problem still leave the most important problems untouched.
Corner identifying, light changing, cross-walk identifying, speed limits, addresses. Not to mention, if the car can't mistake a hunched over person on the side of the road as a curb (because the curb identifies itself) it makes everything else easier to identify. More computation can be freed up to identify things that ARE NOT KNOWN if there is something that IS known.
I mean, I described the reason. The other thing is that perception the way you’re thinking about it isn’t a huge problem for AVs. They’re generally able to recognize all of the objects and elements they need to from their onboard sensors. The problem is a lack of deeper understanding of what they’re seeing, not a lack of viewing angles or range. And more range/angles won’t fix that.
>which will give them at least, a $25 billion market cap.
Uber alone has a market cap nearly 3x that and they have explicitly said they will only become profitable as an autonomous taxi co. Tesla is worth nearly a Trillion on its own. Where do you get 25B from?
It's not neglected, it's just it's basically step zero. Figuring out where the car is in the road is so far down the list of problems a self-driving car needs to solve that if you can't solve it without external aid there's no hope whatsoever you're going to solve the rest of it (like where everything else on the road is and what it's going to do next, unless your plan is everyone who goes out for a walk needs to wear a transponder and indicate as well). What automonous car companies want from the DOT is clearer signage and road markings, and as an added bonus, this also makes human drivers safer as well (but it's not sexy tech, so it tends not to happen).
Most issues are due to sensing humans/new human behavior, not localization. It would be difficult to make a universal system of tags/reference markers that comes close to a decent HD maps solution, even one out of date.
It's a bit frustrating because if all calls were mostly autonomous we could likely have less traffic deaths within a few years. Self driving cars are already great at making snap judgements. Better than most humans. They also make a lot less stupid mistakes than humans do.
The edge cases are the hard part. A human knows what to do when a random construction worker or cop is directing traffic. Self driving car? Not so much.
Kyle Vogt, who co-founded autonomous driving company Cruise in 2013, has been named CEO. He has served as as interim CEO since December 2021 when Dan Ammann, who had been CEO since 2018, abruptly left Cruise.
Generally asking. How does something like this get top three spot on HN? Submitted 30 mins ago and first comment 4 mins ago by a user whose only submission is this. Just curious on the mechanics.
Lots of people upvoting it in a short period of time and nobody flagging it?
The HN ranking algorithm seems to put a lot of weight into both velocity and flags.
A few factors that may have caught people's attention to upvote: it's about an HN company, the author is an HN regular, everything you need to know is in the title, it's written in first-person, and it's a bit unusual for someone to return as CEO.
If i recall correctly the Hacker News ranking gives new items a bit of a boost... possibly more-so for some random users, in order to give them a chance at getting some traction. You might be able to find out more here:
I think that ycombinator sometimes forces up their own links, but that tends to be more obviously related to them like hiring staff to various startups.
They should probably add the FTC ad-disclosure tag to those job posts since the companies are financially related. Or maybe they don't have to since they are related?
What's common in these names - Steve (Apple), Satya, and Kyle?
They all rejected the job when it was first offered to them. It should be a case study and research topic. I think the takeaway here is - don't accept the CEO job right away; conditionally, you should be inside the company. At least that was the case with the above individuals.
Huge fan of Kyle and the Cruise team. I remember covering Cruise in a class assignment back in 2014, and was very skeptical of the idea of selling self driving kits to existing car owners. Ever since, they've gotten acquired (or something) and are holding up their own against behemoths in the space. Cruise's reviews on Blind are pretty bad, something that keeps me away from joining them as an employee, but I suppose chaos and uncertainty is the price one pays for the ambition to achieve something as wild.
As someone who recently left, many of the complaints on blind aren't anywhere close to my personal experiences. I enjoyed the technical side (which is a bit chaotic), but the admin side is borderline antagonistic to the rest of the company. Moreover, the company is on the losing side of a power struggle with GM. The former CEO was a victim of that, seemingly as was the equity and compensation plan.
Can you elaborate on the power aspect - specifically - I thought GM had a minor stake and that Kyle/former CEO etc. have controlling stake? If not, I can see that GM would run roughshod over Cruise's plans. As an aside, Cruise comp numbers on levels.fyi seem pretty reasonable so I am unsure where comp is an issue? Does it relate to refreshers?
I am not as familiar with comma's product, but in 2014, self driving tech was out there enough as an idea for startups to invest in (ie Google moonshot level dreamy), let alone build kits that would retrofit into existing cars and actually work. Thanks for sharing about comma.ai!
Only in the tech industry is it acceptable for a CEO to unilaterally make decrees about what kind of sensor to use and not provide a shred of evidence, no strategy or engineering rigor to support the decision.
I’m talking about Tesla. Imagine if the Boeing CEO one day woke up and started dictating the sensor modality of aircraft to his engineers? People would think it was crazy. Because, it is.
You know, I think the Vision vs LIDAR is very narrow and won't be very relevant in the long-term. If you imagine the full robotaxi stack: cars, training, simulations, back testing, data sourcing, perception, driving decisions, etc, etc, the LIDAR vs Vision concerns only one small part of the stack.
Critically, LIDAR still has to be trained in the same way cameras have to - it'll give you depth information, but you still have to make it accurate decide whether some 3D blob is a person, sign, etc. LIDAR does not help you recognize street markings, interpret signs etc.
That is the big question is: what is cheaper/faster: cheap sensors and more training, or more expensive sensors and less training. As LIDAR and training both get cheaper, it might not end up making much of a difference in the end. Companies will be successful based on how well they execute on the full stack and a slight edge in one area might not be enough to overcome problems elsewhere.
That’s sort of like saying “the touchscreen is only one small part of the iPhone. The OS, the native apps, the camera, the App Store, app developers, hardware accessories, repair service are all part of the stack.”
You’re right, but that “small part” is absolutely critical to get right. It’s what allows the rest of the stack to be built; without that cornerstone, there’s no product.
There had been many attempts at touch controls for computers before the iPhone. Other smartphones were trying. None of them were remotely close to what Apple achieved with their first capacitive screen, and that began a massive change in the way people interacted with their device, and changed the world.
It’s like saying your roof is only “a small part of” your house. You’re right, on some level, but without that roof, there’s not much point in a really nice kitchen.
I think that infrastructure changes and car2car communication, will also make LIDAR less necessary. Imagine if a stopped emergency vehicle could communicate to all cars, AV or not, to slow down, stop or stay out of certain lanes of traffic.
Prediction: This is a prelude to becoming the CEO of GM. How well it works is important to the future of the company.
From what I've read (Farhad Manjoo's article in the NYTimes about the new Escalade), it seems like it's fairly impressive.
The ADAS system in the Escalade has nothing to do with Cruise (confusingly). It’s developed by an in-house ADAS team at GM. Unlikely Kyle wants to become CEO of GM. Until recently the message seemed to be that Cruise was going to operate as it’s own company.
Can someone help me understand why FLIR cameras are not part of the sensor set? Particularly addressing the issue of human related risk? Distance? Target data?
Cruise has been a disaster inside GM for awhile now with a lack of progress to bring Super-Cruise to all models all across the US.
They have had several tenders for the data required to build out the Super-Cruise maps but these have all been a disaster as clear the critical knowledge is gone.
From what I know, GM Super Cruise is unrelated to Cruise LLC. GM has been developing Super Cruise on their own since 2013, and the system relies on a forward-facing camera and 3 radars (front and blind spots) to navigate the road. It links the camera's lane recognition data to 3D road maps stored onboard to anticipate the road ahead. It's also restricted to limited-access roads where pedestrians aren't a concern.
Cruise AVs are much different. They appear to use LIDAR, radar, and camera data to understand their surroundings including intersections, pedestrians, and road hazards.
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).