IMHO a valuable advantage to self driving cars is getting impaired people home safely and legally. If they can do that, it could be a game changer.
Last edited:
Self Driving Car thread
- Thread starter ChaoticUnreal
- Start date
The more I think about that particular problem, the less I understand how I could ever think that an autonomous vehicle would be better at this than a vehicle driven by a taxi driver or transit. The solution has always been there and there is basically no upside to an autonomous car in any of the practical implementations we've seen so far.
Computer driven cars seem like they'd be less likely to murder/drug/rape women (or men!) especially in various states of intoxication.
It's not fun to think about but it happens more than it should.
It's not fun to think about but it happens more than it should.
In my state at least, if you let your kid get drunk and the government finds out, you may be very crossways with The System.
The potential use case is very common in the US. Sober person drives their car somewhere and winds up in a bar, at a friends house, or has several drinks with dinner. While it would make sense for them to be self-aware and not drive themselves home, their decision making is often along the lines of:
- there is no mass transit anywhere nearby, because USA
- uber/lyft/taxi are not affordable, due to surge pricing or raw distance for example
- retrieving your car the next day when you need to get to work at 8:00AM the next day is a huge pain in the ass
- everyone else is just driving home in their cars like normal
- I'm not that drunk, it'll be fine
That's just for people who drink. Other people may not be able to drive due to age or medical issues, and anything that helps people safely adapt to the car culture adds significant value.
I will add that what I see from auto manufactures is not there yet. A robotaxi doesn't solve pricing or car-next-day issues, and self-driving isn't to the point where you can avoid a DUI by climbing into the back seat and letting the car drive itself home.
You also forgot "uber/lyft/etc" doesn't exist here or it only exists to get me to the party/destination and can't get back.
I live 10 minutes south of Ann Arbor, MI and if you take an Uber 20 minutes west of the city to Chelsea you literally can't get picked back up.
The drinking and driving culture in the Midwest is also not ideal.
I live 10 minutes south of Ann Arbor, MI and if you take an Uber 20 minutes west of the city to Chelsea you literally can't get picked back up.
The drinking and driving culture in the Midwest is also not ideal.
I'm non USA I've only recently bumped into "https://www.ecfr.gov/current/title-49/subtitle-B/chapter-V/part-555" "TEMPORARY EXEMPTION FROM MOTOR VEHICLE SAFETY AND BUMPER STANDARDS". At first glance I can't see mention of this here.
555 seems to be primarily around supporting the development of autonomous vehicles. On first glance it has a limit of 2500 vehicles. Am I over thinking the importance of this and the development of vehicles that aren't "merely" customised standard vehicles? Like Zoox or Tesla Robotaxi?
The bit that I wonder if it matters or not is it actually says "not more than 2,500 exempted vehicles will be sold" (my emphasis). If Zoox or Tesla are not planning on selling the vehicle, does it really apply?
555 seems to be primarily around supporting the development of autonomous vehicles. On first glance it has a limit of 2500 vehicles. Am I over thinking the importance of this and the development of vehicles that aren't "merely" customised standard vehicles? Like Zoox or Tesla Robotaxi?
The bit that I wonder if it matters or not is it actually says "not more than 2,500 exempted vehicles will be sold" (my emphasis). If Zoox or Tesla are not planning on selling the vehicle, does it really apply?
Even in a super dense city state like Hong Kong this can be a problem. Taxis have enough work, good fucking luck getting one out to end of a hiking trail half an hour from the city and ~10 minutes out from nearest village.
All the cabs are running major routes and follow the commute schedule. At least some taxi app let you add a big fat tip but even in a smallish village just minutes off a major route / tunnel I had to add like 20% to get a cab once. Bidded less, no response. Upped it, nope. Okay then here you go, finally one replied, 8 min / 5 km away (Yuen Long -> Kam Tin). Weird, it's not that far away even with its own major metro stop.
Would this be any better with robotaxis? Whoever owns them will want to maximize revenue just like everyone else, so they will eventually start making the same basic decisions as human taxi drivers and services, preferring to serve dense areas, not wanting to deadhead from/to some even slightly out of the way place, demanding large tips to do anything other than pickup a passenger where they already are, and drop off where some other passenger is already waiting.
One main idea with robotaxi is the whole-system design to minimize the per-mile cost; most of that comes from it being a BEV and not from the fact that it is autonomous. Tesla Cybercab aiming for $0.20/mi (vs $0.72/mi for Model Y).
With no human in the loop you have a much lower cost structure. You can flood the whole area with robotaxis and not just the city. The cause now is a high minimum price due to paying a living wage.
But if that's all you think about, think slightly bigger. Self driving vehicles can be larger and do non point to point pickups and replace a whole bunch of vehicles with one van. On major routes like airport - anywhere this is already common with shared taxis.
Cities, villages already pay major sum to run public transport. They could subsidize a few shared taxis and base them in a smaller town or village in addition to the bag of money paid to run a regular bus service with way too few travellers. It might even replace it one day, switching from fixed route to point to point (with detours to keep it affordable).
https://www.theverge.com/news/848843/waymo-san-francisco-power-outage
I guess, fail safe, but this doesn't seem great.
I guess, fail safe, but this doesn't seem great.
I enjoy that any viable public transit system eventually boils down to existing public transit
The big problem with any personalized transit system is that you need to design public transit to a public use standard. People are not going to accept driving in a cloth seated self-driving car where all the seats are ripped and barf is on the ground. Cleaning and maintaining cars is a relatively expensive affair, on top of the quite high costs of cars. That's why public transit is designed the way it is: bright lights, social control through having multiple people share a vehicle, easily cleanable, durable surfaces and seats, etc.. And the per-person cost of a transit vehicle therefore ends up being quite low.
And then if you can put a lot of people on a single vehicle, you suddenly don't have significant driver costs anymore, almost all the cost is in the vehicle and exploitation, and those are significantly cheaper than individual cars.
And then for the last mile, you can have people use small individualized transit vehicles, they can be roughly person-sized, maybe even with just 2 wheels - large ones for comfort. Maybe self-propelled with a little electric motor to assist in bad weather. Like a bike! Oh hey, we also have this entire transit system fully figured out 20 years ago.
No one is trying to replace full busses or double-deckers with robotaxis. But to say, we have bikes, why bother with something fancier - that just ignores reality.
If I can call a robocab as easy as driving, even at +50% cost per km, I wouldn't bother to have a car.
Tourists don't bring their own bikes. You don't bring your bike all the way from home, in the train, to the destination to bike a few km there. Rent a bike / scooter is fine for some, useless for someone with a suitcase and not available outside of major stations - so useless for 95% of metro stops and most train stations.
The "solution" now is a dumb bus route starting and ending at the same train/metro station that is always empty running 3 times per hour, sets a
for the city/village that public transport exists and covers most addresses, but in reality is the worst of all options and only used by people who can't afford Uber or have someone to pick them up from the station.
If I can call a robocab as easy as driving, even at +50% cost per km, I wouldn't bother to have a car.
Tourists don't bring their own bikes. You don't bring your bike all the way from home, in the train, to the destination to bike a few km there. Rent a bike / scooter is fine for some, useless for someone with a suitcase and not available outside of major stations - so useless for 95% of metro stops and most train stations.
The "solution" now is a dumb bus route starting and ending at the same train/metro station that is always empty running 3 times per hour, sets a
for the city/village that public transport exists and covers most addresses, but in reality is the worst of all options and only used by people who can't afford Uber or have someone to pick them up from the station.I feel like you're new to the world of transit arguments. All of this is extremely well-trodden ground.
Yes, there is a place for taxis. In the middle of the night, in bad weather and if you need to go to some hermit well out into the countryside. All of this is an extremely limited use case and inherently makes this a small, expensive market. This is not generalizable. The per-unit cost of transit is untenable in car form. There is no way to make cars cheap, reliable and acceptable enough as transit and by god have people tried over the years.
Driverless cars also already exist as transit, and have for close to 40 (!) years now! In the Netherlands, even. Right now they're called Greenwheels. They are considerably cheaper than an Uber, yet absolutely nobody uses them. In almost all instances where they've been tried in a subsidized manner, they have vastly underperformed their expectations. Even as BrengMij, a service that put cars at transit hubs and allowed you to drive it to your house and had somebody pick it up after so you didn't need to bring it back, it completely failed. Because despite what people think robotaxis solve, they can't ever achieve the scale necessary to make them affordable no matter how few drivers they have.
Anyway, as I said, this is all well-trodden ground. Robotaxis don't solve transit, they just don't. They're a small cost saving on larger taxi companies, that's really all they can ever be. And a political tool to prevent proper transit to be realized anywhere, as you already allude to.
Doesn't make them uninteresting from a technical/science perspective though. I love talking about SDCs, and I have for the past 15 years that they've not-existed. Almost nothing in these discussions has changed at all, try opening an SDC thread from 2017. Exactly the same discussions going on.
Yes, there is a place for taxis. In the middle of the night, in bad weather and if you need to go to some hermit well out into the countryside. All of this is an extremely limited use case and inherently makes this a small, expensive market. This is not generalizable. The per-unit cost of transit is untenable in car form. There is no way to make cars cheap, reliable and acceptable enough as transit and by god have people tried over the years.
Driverless cars also already exist as transit, and have for close to 40 (!) years now! In the Netherlands, even. Right now they're called Greenwheels. They are considerably cheaper than an Uber, yet absolutely nobody uses them. In almost all instances where they've been tried in a subsidized manner, they have vastly underperformed their expectations. Even as BrengMij, a service that put cars at transit hubs and allowed you to drive it to your house and had somebody pick it up after so you didn't need to bring it back, it completely failed. Because despite what people think robotaxis solve, they can't ever achieve the scale necessary to make them affordable no matter how few drivers they have.
Anyway, as I said, this is all well-trodden ground. Robotaxis don't solve transit, they just don't. They're a small cost saving on larger taxi companies, that's really all they can ever be. And a political tool to prevent proper transit to be realized anywhere, as you already allude to.
Doesn't make them uninteresting from a technical/science perspective though. I love talking about SDCs, and I have for the past 15 years that they've not-existed. Almost nothing in these discussions has changed at all, try opening an SDC thread from 2017. Exactly the same discussions going on.
I have a subscription on them. It's often available, but you have to hail them at their resting place and return them there. Not usable as last mile solution. Total different thing from SDC.
And what evidence do you have for that? The old shit you mentioned are not taxis and come with severe restrictions. Nothing like getting a cab in Osaka, Kyoto and Hong Kong, something I did many times past few weeks and like, every year or so. And the taxi rides to and from Schiphol.
That's what SDC is supposed to bring, cheaper taxis as the basis, pooled or private.
I feel like you don't get out often enough. Netherlands isn't the only country and Randstad not the only region. There are plenty of times I wished public transit is better in Austria, Germany, anywhere without frequent service and the ridership forcing many to drive.
SDC won't replace well filled busses, never mind trains, trams and metros. But busses running mostly empty will be endangered species and taxis asking for €3/km are the first to go. Make it €1 and widely available, easily hailed and it will be the first choice for a significant group. Bike - train - taxi is then a valid trip suggestion, instead of bike - train - fucked.
The relatively affordable cabs in Asia show how it could be, very few private cars on the road and a mix of absolutely all modes of transportation between bulk carriers (metro, trains, double decker bus) to more flexible modes (mini bus, shared/pooled taxi) to the ultimate luxury, taxi. In the Netherlands we have quite a few holes to fill, busses running a loop around stations causing a 15 minutes drive to explode into bus-metro-bus route of an hour total time. If you don't travel along the suburb - city backbone you are screwed with those stupid loops on both ends of the trip.
It can also enable services that are less hub-centric. Compare that with airlines with a hub and spoke design, like the middle eastern flag carriers operating A380s, vs tons of direct flights using smaller planes, at the extreme end, Intercontinental with A321neo LR/XLR narrowbody. It shortens total trip time dramatically.
Aight dude, just go read up on transit. You're using literally every ignorant argument in the book that shows you just never thought about transit in depth. You're not adding anything here and just embarrassing yourself.
Depending on how you enjoy your informational resources, there's amazing youtube channels on the subject like RMTransit, City Beautiful, even combative annoying nerds like Not Just Bikes and Adam Something you may enjoy (if you like arguing on the internet among nerds like I do).
There's the urbanist blog as a good central reading resource. There's urbanist twitter if you feel like losing every single brain cell you have left.
I hate quoting myself, but I feel like that I've went over the economics argument of SDCs, as it pertains to cabs, quite extensively earlier this year.
As for driverless cars, dockless versions exist, and I have used them extensively in several European cities. The economics to make them work are hard, as can by seen that first DriveNow (Merc/Daimler), DriveNow (BMW) merged into ShareNow, and then got bought up by Free2Move (Stellantis).
This type of consolidation happening despite the fact that all involved companies are/were OEM-owned and hence likely able to purchase their vehicles not at commercial market rates but near at-marginal-cost of the OEM owning them is not indicative of an industry with healthy profit margins.
Others have chosen ... interesting business models that seem to essentially exploit the sort of economics you only have if you're already an established rental car company (like Sixt), or where the business model might essentially burn down to renting out the car as a loss leader to being able to issue dodgy 'fines' for supposed wrong doings like smoking of their customers (the company that starts with M and has ... interesting opinions on how contractual penalties work)
As for driverless cars, dockless versions exist, and I have used them extensively in several European cities. The economics to make them work are hard, as can by seen that first DriveNow (Merc/Daimler), DriveNow (BMW) merged into ShareNow, and then got bought up by Free2Move (Stellantis).
This type of consolidation happening despite the fact that all involved companies are/were OEM-owned and hence likely able to purchase their vehicles not at commercial market rates but near at-marginal-cost of the OEM owning them is not indicative of an industry with healthy profit margins.
Others have chosen ... interesting business models that seem to essentially exploit the sort of economics you only have if you're already an established rental car company (like Sixt), or where the business model might essentially burn down to renting out the car as a loss leader to being able to issue dodgy 'fines' for supposed wrong doings like smoking of their customers (the company that starts with M and has ... interesting opinions on how contractual penalties work)
Last edited:
It's probably worth challenging the assertion that the driver is even the expensive part here. I mean the best thing Waymo could do from a gross margin perspective is switch to conventional electric cabs with drivers. In the end, automating the driving maybe cuts costs by 20% in markets with expensive labor assuming you don't need to hire rooms full of people that are able to intervene/assist in the event that there's a problem and by a fraction of that in low labor cost markets, again assuming the system is perfect and no additional labor is needed to track the autonomous fleet.
Lets be clear here. Waymo is never going to amortize the cost of that expensive sensor package. It's a money sink. And there just isn't revolutionary cost savings to be had by implementing self driving cars because cabs are already cheap.
Lets be clear here. Waymo is never going to amortize the cost of that expensive sensor package. It's a money sink. And there just isn't revolutionary cost savings to be had by implementing self driving cars because cabs are already cheap.
But why? Assuming you write them down to zero in 4 years, the estimated cost of $150000 is only $9 per day. That's not a huge number to overcome, just one extra short trip per day covers it already.
Pff, you're always in your high castle looking down on plebs. Make a valid argument for once instead of just sprouting unsupported claims.
The model I am looking at exists in real life in multiple countries and cities. That it isn't well described by western youtubers / urban planners doesn't mean it doesn't work.
And I do follow most of those on YouTube. They have a fetish for Dutch and Europe. It's great here, but there are plenty of issues with our transport system solved better in the far east. I have named them.
Stop with those information-free claims and start reacting to actual points. Here are the main claims I made:
1. Self driving cars, if they can drive just as well as humans, will replace labor in taxis.
2. Netherlands needs better point to point transport. Better than those stupid loops around metro and train stations. SDVans modeled after public red/green minibus and UberPool can help. Non self driving version exists, and without huge subsidies are not affordable - in NL it is now only for people with mobility issues.
3. Netherlands has too many private cars. Many cities in Asia has way less, in places where taxis are not luxury only for the 1%. Look at any busy street, excluding cargo transport vans and trucks, and often most if not 75% of vehicles on the road are public transit. Bus, minibus, taxis. Not >95% private cars like here in Rotterdam. SDTaxi and minivans, if they can work out point 1, sensors and software, can help push cost down and start shifting cars on the road towards public transit.
And a new one for those with a bike fetish:
4. Bikes are great if you can reach your destination without transfers. They are great to the local station. They are useless for the last mile on the far end - not enough bike rental (OV fiets) available so you depend on the slow option, usually a low frequency bus.
They are also not great in weather - as soon as it gets colder and wet, traffic on the road increases even though everyone knows you will be stuck in horrible traffic. This effect has been quantified: https://www.researchgate.net/public...choice_Empirical_evidence_for_the_Netherlands and https://uni-koeln.de/en/university/...utch-cycle-twice-as-much-as-germans-in-winter - and we are diehards compared to other nations at sticking with the bike. Trips below 5km are still fine, but trips over 5km are at risk for mode change.
No amount of bike path and other infrastructure can change this. E-bikes though helps, it increased winter cycling by making it less prone to wind, just turn assist knob to turbo, I did that with headwind on the Zeelandbrug / N256, towering over open water and extremely windy. The three effects: temperature, rain, wind, 2/3 are unfixable but the last can be with €.
It also requires people to adjust though and that might be the hardest part. We laugh that videos for tourists coming to Europe encourage people to walk more to prepare themselves, just taking public transit and walking 5-15k steps between the trips a day is a huge step up from their normal lifestyle, and this lifestyle exists here too. Not everyone can and will bike, and the other option shouldn't be "too bad, well your next bus is in 25 minutes". Or just drive, car modal share is over 40% here. I rather have those take a (robo)cab.
People who can bike already do so. Convincing the other part of the population, cycling has a modal share of 20-25% here, will be hard to impossible. It's not the solution for everything.
I am not saying that we can / should beat the end boss city state with only 1.5% private car share, but 40% here is too much and we can do better. But not much with the existing stiff and ancient "hub and spoke" setup. There's a reason public transit only captures ~20% of the trips even in a very well done setup in Rotterdam, NL average is dramatically lower at ~12%. Many of my trips past week double to triple in time compared to driving, so that's a nope.
Last edited:
How is it that Chinese cars have multiple lidars in cars under $30k or even $25k?
And they don't sit in those giant housings like the Waymo cars have.
Waymo can't source these units? Also, these Jaguar iPace models are old now in EV time. I heard they've developed another platform but hasn't deployed as widely.
iPace was like $75k when it came out. Google probably got a deal for buying so many but now, the sticker on many EV SUVs are much less.
And they don't sit in those giant housings like the Waymo cars have.
Waymo can't source these units? Also, these Jaguar iPace models are old now in EV time. I heard they've developed another platform but hasn't deployed as widely.
iPace was like $75k when it came out. Google probably got a deal for buying so many but now, the sticker on many EV SUVs are much less.
The government subsidizes the shit out of the company for both behind the scene control and (amusingly enough) to keep foreign products/influence out as much as possible.
If the US government simply gave Ford 75% of the cost of an EV for "nothing" in return they'd make super cheap cars too.
The current Waymo jaguar platform is already legacy, Jaguar stopped producing the cars and Waymo still has around ~1k they need to retrofit and use until they die. Here are the pictures from the retrofit factory where you can see (all?) the remaining jaguars waiting in line for retrofit:
View: https://www.reddit.com/r/waymo/comments/1prp9a6/waymo_factory_december_2025_update/
As a comparison point, a Tesla factory can produce about 1 Model Y per minute.
The next Waymo generations are already up-and-coming, you've seen the hyundai ioniq5 based ones and also the zeekr van-like one, here is a video from 1 year ago:
View: https://www.reddit.com/r/waymo/comments/1hxudin/detailed_look_at_the_6th_gen_waymo_driver_at_ces/
I see the zeekr ones on the streets around here all the time.
View: https://www.reddit.com/r/waymo/comments/1prp9a6/waymo_factory_december_2025_update/
As a comparison point, a Tesla factory can produce about 1 Model Y per minute.
The next Waymo generations are already up-and-coming, you've seen the hyundai ioniq5 based ones and also the zeekr van-like one, here is a video from 1 year ago:
View: https://www.reddit.com/r/waymo/comments/1hxudin/detailed_look_at_the_6th_gen_waymo_driver_at_ces/
I see the zeekr ones on the streets around here all the time.
In San Diego, where Waymo is set to launch next year, I generally see the Zeekr vans driving around right now.
A LIDAR is a diode laser, an array of spads, some mirrors and simple plastic lenses turned by a motor. The components are extremely inexpensive (spads cost literally dollars) in volume, but require huge one-off engineering costs (I.e. software/firmware). Engineers are cheap in China, and one-off costs become insignificant if spread across enough units.
People in tech who don't do imaging had this idea that some how a brushless motor and a spinning silicon chip was somehow going to be impossibly expensive, but they were dead wrong.
Perhaps the previously expensive part was more in assembling them to consistent quality, and calibrating them to sufficient tolerances? Like, sure, anyone could source those parts and throw something together, but that doesn't do much good for a system receiving the output if it's trash.
There was a story about a US lidar startup whose founder was a very young guy, so he was looking like some wunderkind because he'd be able to produce lidars much more cheaply.
But that may have been before Tesla decided they would only use vision and Waymo seems to be rolling their own or committed to this older design where you can see the spinning components inside the cylinders protruding from the Waymo cars.
But that may have been before Tesla decided they would only use vision and Waymo seems to be rolling their own or committed to this older design where you can see the spinning components inside the cylinders protruding from the Waymo cars.
The expensive part is probably patent licensing. Thus how cheaply Chinese companies can produce them
Due to manufacturing reasons, SPADs are physically huge compared to CMOS pixels (10s to 1000s of microns vs single microns), so tolerances are really, really forgiving vs something like a cell phone camera. For example, if you're trying to focus on a 200 micron wide pixel, you can actually be off by a lot before it is going to make any optically detectable difference, and some SPADs are 1 mm wide or even more. So from an assembly/quality perspective, your tolerances are very forgiving by optical standards or compared to something like a car backup camera just because the components are physically huge and the number of pixels is very, very low.
Similarly calibration is very reasonable. Your ranging is set by the speed of light and the delay of CMOS of transistors, both of which are highly predictable. Your rotational position is determined by a position encoder on the motor, but typical resolutions are ~ 0.1 degrees, so your encoder only needs to have a few thousand points per revolution, and probably doesn't need calibration. Accuracy in elevation is the one that is most sensitive to alignment, but it is also the axis you care about least, so you don't even need crazy high precision machining.
Not that engineering is trivial, making something that will work after bouncing around in a car for 100,000 miles and in freezing cold weather and scotching hot Arizona sun is never trivial, but once you have the engineering worked out, things like electronics, simple lens assemblies, DC motors, etc. are highly manufacturable at low cost. The rotational kind of LIDAR is really nice because everything is on axis and you're spinning the optics rather than trying to precisely focus onto millions of microscopic pixels. Having 128 pixels instead of 10,000,000 just makes everything easier.
----
LIDAR trips a lot of tech people up because it is a simple hardware solution to a difficult software problem. It is absolute possible, with enough trillions of multiply adds per second and precision optics to figure out how far away things are from parallax, context, etc. Incredibly hard and expensive, but possible. Or you could bounce a laser diode off it, measure that distance using time of flight, and get the same thing more accurately from something with the CPU power of a budget smartphone. People's instinct is that it is always better to bet on software algorithms over fixed function hardware, and maybe in the very long term that might be true, but not today.
Last edited:
I doubt the devices are actually expensive, even in the USA. When you look at low-volume precision optical devices, the price might be 20, 50, even 100 times the cost of the hardware because the tiny number of units sold must pay off the huge engineering costs that went into designing it. But if you're Waymo or Baidu and you're making thousands of units per year, you don't care about the sunk costs that went into designing your LIDAR, you care about how much it costs you to make the units you need for the coming quarter's expansion. That number will be easily 20 times less than the list price of a LIDAR system.
LIDARs have the exact same manufacturing problem that chipscale modems had for a long time: processors and transistors weren't fast enough to do anything active to the analog signals, so you needed very large physical structures to do filtering, tuning, etc.. Modems suddenly became trivially cheap to make when we could insert active electronics into the frontend, because all the 'black magic' physical structure disappeared.
LIDAR is in that same transition. Remember those giant hybrid ceramic substrates for radar waveguides? LIDAR needed the same stuff and even larger than commodity radars. Those went away, and then we still had the precision rotating/oscillating laser assemblies. Those are now also being replaced with MEMS mirrors similar to DLP, although still requiring precision optics. The last step in the cheapening process is getting rid of the complexity in optics, which is where current acquisitions in the space are partially focused. Ultimately, there is no real price floor to LIDAR, the technology could in principle be commoditized to prices well below $100 for a ~100kpt/s unit (something that right now is still an order of magnitude more expensive).
By the way, this specifically pertains to automotive lidar, there's whole separate constraints on e.g. satellite/airborne LIDAR and telescope LIDAR. The required beam power is so much higher that the laser presents a high cost floor.
Exactly how fast do you think the analog bandwidth is on a SPAD array used in automotive lidar?
I really don't know anything about the specifics of the AFE of these things, just their construction and how they evolved over the past ~15 years or so. Moving away from hybrid construction is the primary driving force of cost reduction in 'early' LIDARs (i.e. going from ~$10k to $1k units).
Few hundred MHz for 1 cm resolution.
I'm curious when you think it will be possible to make processors and transistors fast enough to operate at MHz frequencies?
Last edited:
Don't really get what you're going at. I understand the mixing frequency is going to be easily slow enough, but the problem is with - effectively - your carrier frequencies, right? That's where you're doing the filtering, and for lidar that's e.g. 905nm/360GHz (or with lidar it's usually a wave package of order 10-100 wavelengths, so 3-30GHz). So to make e.g. a mixer or narrowband filter, traditionally you do that with a tuned cavity or with passive elements on a waveguide. That's not something you can do on FR4 with discrete components because it's too inconsistent, and it's not something you were able to do in silicon until pretty recently because they didn't run in the GHz range and also had pretty hard limits on the kinds of passives you could create on silicon.
That's what's different now. It is now possible to mix, filter, multiplex, create oscillators, do all that stuff in the tens of GHz directly on silicon, and with plenty of power too.
I'm pointing out that if you look at the bandwidth of these devices you'll see that they're orders of magnitude slower then you're assuming because they work in a different way than you think.
Actual automotive LIDAR is incoherent and so does not have a mixer. You simply pulse out light and time the reflection. Your detection electronics are essentially a SPAD and a timer triggered by the rising edge from the SPAD. It is really simple compared to what you are thinking and why I keep saying that these systems can be made really cheap in volume.
You are thinking of coherent lidar, which is not used in automotive. Mixers at optical frequencies are implemented with interferometry using optical fibers or silicon waveguides rather than transistors or filters. This down converts the THz optical frequencies to whatever bandwidth you want to work at, which is set by the tuning rate of the laser source driving the system. There is a similar way to do it with pulsed rather than swept lasers, but it is less practical. If you don't mind reading out a few thousand points per second, you can tune the laser slowly, read out at kilohertz bandwidth and detect on your phone's audio jack. Typically though you operate at few hundred MHz, which gives you a few hundred megavoxel/s rate, and is kind of a sweet spot for inexpensive A/D chips and FPGAs. Plus at a certain point you just run out of photons if you try to scan faster, especially if you want the laser to be eye safe.
Coherent detection has advantages (immunity to interference, potential for wavelength-scale resolution, doppler information), but it is not really clear that any of this is needed for automotive applications or worth the increase in complexity. Perhaps silicon photonics will eventually become so cheap that it'll be adopted, but that is probably decades away and it is clear that mass-adoption of SDCs will not be waiting for that to happen.
I was reading a reddit post or something where someone said the old Uber ATG systems were a pile of nvidia gtx 1080s in the trunk.
From what I understand today, the lidar is cheap and easy enough, and now the processing and computing hardware is also commodified enough (between nvidia and tesla and whoever) so now the secret sauce is all in the software. Who has the best lidar -> driving software stack?
From what I understand today, the lidar is cheap and easy enough, and now the processing and computing hardware is also commodified enough (between nvidia and tesla and whoever) so now the secret sauce is all in the software. Who has the best lidar -> driving software stack?