In a document prepared for a Korean trade delegation to Silicon Valley in May, and obtained by IEEE Spectrum, Strobe claimed that its prototype lidar had a range of 300 meters, a processing time of fewer than 45 milliseconds, and cost less than $100. It also noted that its first product would be commercially available in the spring of 2018.
Probably fabrication cost, given that it's a prototype. It's a research project so their estimates are a bit pie in the sky right now. Until it's mass produced it'll be expensive, but who knows, maybe GM will pull it off.
Notable that this is an FMCW system. Very few commercial players use frequency modulation, they all use either AM or pulsed lasers.
Rep rate is not so good. 45 ms = 20Hz maximum. That's poor compared to Velodyne (15kHz or so per transceiver) or surveying stations which can push 1MHz. Even hobbyist systems like the Scanse manage 4kHz. I'd guess that the problem is, in part, because you need to chirp for a certain amount of time. Pulsed systems often use avalanche diodes that trigger from relatively few photons.
It's 1D, so you've still got to scan it which is bad for robustness. I'm more excited for solid state LIDAR that uses beam shaping (or however they're doing it) to scan without moving parts.
The range is good though, FMCW is great for that (and the accuracy should be good too). The main advantage of FMCW is that you can still get the high accuracy of AMCW systems (i.e. at a lower bandwith than direct time of flight), but you don't have the phase ambiguity problem that AM does.
If that's their per-sample time, they probably would've dissolved by now after admitting total defeat - 20 Hz is totally unusable for navigation. I'd guess that that's something more like maximum latency from real world to computer, either due to deep pipelining or because they communicate data frames rather than a stream of individual samples. Very easy for that kind of thing to fall through when talking to the media.
> Number of pixels and field of view is controllable in real time to 'zoom' the resolution.
and the IEEE quote is a bit ambiguous (emphasis mine):
> An image on Strobe’s website seems to suggest that the lidar will be much smaller too, with a footprint as small as an American quarter. A number of these would be needed to generate the 360-degree coverage provided by the rotating, roof-mounted lidars seen on many experimental self-driving cars today.
Needing several implies a fixed field of view, but that could be either 1D or 2D. Given the size of the device (a US quarter apparently) it could be an area imaging device, but that would imply a pretty big improvement in optical technology to get enough signal-to-noise per pixel. I know they claim FMCW requires less sensitive detectors, but there's a limit.
"pixels" and "field of view" along with the comparatively terrible "20 Hz" rate strongly hints to me (a total layman, admittedly) that these are full 2D frames at 20 Hz, not single pixel samples at 20 Hz. The thought that it would take a full 45ms to sample a single pixel sounds absurd, even for research tech in this area when industry standard is pushing 100 kHz. A 2D surface sampling, say, 200x200 = 20k pixels per frame * 20Hz sounds much more reasonable.
If it was only 1D, it would be utterly unfit for navigation if you just scatter a few of them around on the vehicle in fixed locations like their marketing suggests. You can't replace a spinning roof-mounted lidar with just a few fixed 1D sensors. The spinning lidar gets something on the order of one sample per degree, right? To even come close to replacing that with fixed 1D sensors you'd need on the order of 360 of them around the vehicle, and their marketing doesn't sound like that they expect you to use anywhere near that many.
What? I've seen lots of autonomous robots that operate at 10 Hz, including mission critical systems for the military. Compared to cameras, which might operate as fast as 60 Hz (but are far more likely to operate at 30 Hz or a fraction thereof) I am confused where you're getting these numbers from. Do you have any citations?
20Hz refers to a single point measurement, as this is (to the best of my knowledge) a 1D LIDAR. Every 45 ms you make a single range measurement. You still need to scan the sensor head to image your scene. Though as the parent comment mentions, it may have gotten lost in the noise somewhere. There's no way GM is using this as a 1D sensor on a car.
the light cone doesn't have to be focused, a single scan can illuminate a big area, it should be enough that the phase is coherent over the area. or so I'd imagine, but on the other hand, a 100ms ping doesn't mean 10*packetsize throughput per second on the internet either.
IMHO it's fun to think about the economics of autonomous vehicles. For example, here's a neat article on "second order effects" of electric and autonomous cars [1]. (examples: Increased use of autonomous cars could mean less space lost to parking lots in cities. Higher utilization for cars could mean faster turnover (~3 years/car instead of ~10).)
One thing to think about: autonomous cars might make a big difference for car-sharing services or trucking, but they're less essential when you're driving the car yourself and parking in your garage. I think traditional cars will be around for a while. (Particularly in lower-density areas where there's less demand for ride-sharing services.) This article [2] suggests regular cars may be around until 2050 or so:
"According to IHS, a firm that provides automotive forecasts and insights, sales of autonomous cars, including driver control, will begin
by 2025 and could reach 11.8 million in 2035; sometime after 2050, says IHS, almost all vehicles will be
autonomous."
So your hobby is probably good for another 30 years at least?
Can you imagine how much additional livable space is going to be available across the world if people convert their garages into another room?
Assuming 12 ft by 22 ft (minimum space for a one car garage, rounding down nation wide to be conservative), that's 264 sq ft. In 2016, there were about 125.82 million households in the United States. That's 33.216 billion sq ft of additional living space!
They say they're not making more land; I'd argue we're going to be way more efficient with the land we have in the future.
You are counting households that live in apartments there.
And lots of garages aren't attached (on this block there are 12 garages with 1 being attached to the house).
Another way to say 30 billion square feet is 15 million to 30 million housing units. A drop in the bucket of the additional housing that will be needed globally in the coming decades.
> You are counting households that live in apartments there.
Good call, my mistake.
> Another way to say 30 billion square feet is 15 million to 30 million housing units. A drop in the bucket of the additional housing that will be needed globally in the coming decades.
I don't think this is the case. Population is on the decline in the industrialized world. Africa, Indian, and China need a whole lot of housing though.
They seem to have killed a bunch of urls, but they do tables of future expected population. The US population is very much expected to exceed 400 million before too long.
Whether detached or attached, once you get out of denser urban areas, land isn't really the limiting factor. In many cases, a garage is a lot different from livable space. I have an attached garage and you'd probably just tear it down to make living space out of it--and that would just result in a bigger house.
I figure a great deal of this will be standardized with modules you swap out like you do a water pump and sensors like light assemblies. plus it is not like all these older cars are going anywhere.
people vastly over estimate how fast self driving cars will arrive and just as much the electric only revolution
GM right now is the most vertically integrated of all the companies making meaningful progress on Robotaxis. They have a dedicated assembly line set up building off the Chevy Bolt platform, and intend to have 'thousands' of them on the road before the end of 2018. They're building their own ride hailing app, called Cruise anywhere, currently only available to GM employees. They've got On-star, which provides in-house expertise with connected car and vehicle diagnostic services. GM's Maven subsidiary offers car sharing services.
But when Robotaxis really proliferate fleet management infrastructure will be very important. Apple and Waymo have made partnered with Hertz and Avis respectively, but General Motors can utilize the real estate and expertise of it's existing network of dealerships, which can save them from a great deal of capital investment as Robotaxis dissemintate.
Before last December there wasn't a whole lot to be known about Cruise's progress, but since then it's just been one reason after another to be getting excited about what they're doing.
>...but General Motors can utilize the real estate and expertise of it's existing network of dealerships, which can save them from a great deal of capital investment as Robotaxis dissemintate.
I was involved in a project at a major US auto maker where they wanted to utilize spare capacity at the dealers on something similar to what you are discussing. In this case, we would pay the dealers to utilize that capacity.
In some states, the dealer networks are large, politically very influential, and are not inclined to allow anything that will disrupt the status quo of selling and leasing cars. They absolutely had their hands around the throat of the auto maker and forced us to go another route just because they saw what we were trying to do could possibly change their income flow in some small way.
In theory, GM has the infrastructure and is making the right acquisitions. In practice the vast majority of the company is entrenched in selling and leasing cars and many people in the company don't have the imagination to see the world differently. GM has a lot of internal and external politics to overcome before they are successful with autonomous cars and Robotaxis.
That article was interesting. I've been a truck driver, with recurring jobs throughout San Francisco, and can attest to how nerve-racking it can be. In order to drive trucks there successfully, you have to know the roads really well, as your GPS is often worthless (so many different rules, changing conditions, unpredictable drivers/cyclists/pedestrians, as well as truck-specific restrictions). From a pedestrian safety perspective, I really appreciate that there are companies that are prioritizing San Francisco as a testing ground.
I still think Seattle is more difficult to drive in. The streets are narrower (most residential streets are two way, but only wide enough for one car width, requiring drivers to negotiate with other drivers on how to proceed). There are far more 3,5,6 and even 7 way intersections[0]. Just as many hills as San Francisco, but the several poorly-intersecting street grids give the hills much worse visibility. And I'm willing to bet we have way more construction than San Francisco.
We haven't even seen a driver-less car floor it on a green to make a left before the opposing traffic moves
Of course, this isn't an issue in San Francisco, because the intersection is completely gridlocked. The only way you're going to make a left turn -- or even a right turn at many intersections at rush hour -- is by violating various traffic laws up to and including the Pauli exclusion principle.
I can't imagine how to build a self-driving car that can perform both safely and effectively in any major urban environment... and I'm not the type of person who says "I can't imagine" lightly. It would literally be easier to work on autonomous rockets at SpaceX. With self-driving cars, the engineering problems are largely social, not technical.
Probably even worse in a place like Manhattan. Tuning the "aggressiveness" setting so that you hit the right balance for a given situation between making forward progress, not blocking the grid, avoiding all the pedestrians who jaywalk (including myself) as soon as there's an opening, etc. seems like a really tough problem.
SF has plenty of unprotected left turns against traffic. I haven't seen a driverless car outside Mountain View, but hopefully they actually make those turns and don't just get stuck?
That said I remember the occasional drive straight down a cliff in Pittsburgh beating even the SF hills.
Aren't they still working out the actual software... To make all of this work...hardware wise they are ahead but people around here have been tooting the GM horn just because they can produce cars.
I never thought about car companies being able to use their dealerships for their fleets. That will actually be a pretty huge advantage in smaller towns where there won't be enough business to have the cars driving continuously.
It's not really "their dealerships". A car dealer is an independently owned franchise. None are owned by the car manufacture.
Of course, the manufacturer can say "we will give you $x off your car purchase if you let us use your lot" but the reality is the dealership can just tell them to take a hike.
Also, manufacturers lease to rental car companies. So that is an option as well. I can imagine Enterprise having unrented cars in a lot that switch to robotaxis...
Vogt did an interview with Forbes earlier this summer, in which he implied that the Lyft investment was done on the initiative of some other arm of General Motors, and that Cruise doesn't need no stinking rideshare partnership. I guess we'll see how that plays out.
Strobe seems to be using some very advanced physics. Here's a paper on quantum LIDAR.[1] The semiconductor technology is exotic.[2] The detectors in the research papers require cooling to superconducting temperatures, but apparently that can be overcome. It's amazing this technology is making it to the automotive environment this fast.
Anyone care to explain the technological situation of Cruise? I feel like I read mix messages all the time about them. Some say that Cruise is just putting up smoke and mirrors while others make it seem like they're the closest thing to level 4/5 autonomy compared to anyone else. I'm super curious where they actually stand in terms of technology.
Heh, that would be inside information would it not? If nothing else, the Waymo v. Uber lawsuit shows how hotly contested self driving technology is, and LIDAR technology in particular.
I've built a number of mobile robots over the years, and watched many others build them as well as part of a robotics club. One truism is that most robotics problems are pretty easy given ideal sensors and sensors for every variable :-). When I started building robots I quickly understood that 'sensor fusion' is more about trying to tease reality out of a very noisy cloud of data than it is AI or fancy software. I spent a month trying to get a robot to go precisely straight so that my dead reckoning algorithm would work.
True, it would be pretty awesome if Cruise had some of their code open sourced to have a peek at their technology. I was looking at Google's Cartographer and it's pretty neat to see some of their SLAM work. It would be nice to see something equivalent from Cruise that shows what they're capable of without revealing all of their tricks.
I don't disagree, but you did see where Levandowski got $120 MILLION DOLLAR BONUS from Google right? There are literally billions of dollars of investment risk riding on the outcome of self driving coding and so nobody in a 'company' context would be likely to be motivated to share anything at all about how they were going about it. It would be like asking folks who wrote lottery software to open source their PRNG implementation so that people could see how cool their technology was :-). The brief upside of 'coolness' is massively outweighed by the major downside of 'used this against you and cost you millions.' right?
The problem is that outsiders aren't going to have a clear picture of the actual state of each implementation.
The CA DMV autonomous vehicle reports of accidents and disconnects are the closest thing available to objective data right now. Waymo is way ahead on disconnects. Cruise is getting rear-ended a lot lately. Google used to get that a lot, but they seem to have gotten past that.
(The main compatibility problem with autonomous vehicles so far is being rear-ended. The typical situation is that the autonomous vehicle advances into an intersection, detects some good reason to stop like cross-traffic previously occluded, stops, and gets rear-ended. The accel/decel profile for entering occluded situations may have to be made more compatible with human behavior to get the human drivers behind to behave properly. This isn't a big problem; damage in those collisions is very low.)
If autonomous vehicles become a success you'll just be able to buy LIDARs for super cheap because the market for them will drive prices down significantly.
You'd think so, but there are a lot of things that have a big commercial market but are basically non-existent in the consumer space unless you're willing to rip them out of commercial products. As just one example, you can buy one 7" HDMI touchscreen (just the screen) for $100 shipped [1] or for the same money you can buy two Amazon Fire tablets with the same size screen but twice the resolution, plus it has a full mobile SOC with a quad core processor and a camera and I mean it's a full tablet. Not just the screen. And it's half the price.
Even if Amazon is losing money on each Fire sold, I can't imagine they're losing that much. They're getting those screens for a hell of a lot cheaper than I can, and better quality too. I imagine LCD touchscreens have a far bigger consumer demand than LIDAR sensors ever will, so the dream of cheap LIDAR sensors from Sparkfun will forever remain a dream. Grabbing one from a junk yard will be the better option.
But those include a "beefy DVI/HDMI decoder", which the tablet doesn't need, since it can dump the raw image directly from the graphics chip. You can probably find a cheaper version if you can do the same.
EDIT: To corroborate my last point, a replacement screen for a Nexus 7 is just $22 (free shipping), and includes digitizer.
Similarly, I'd imagine a LIDAR sensor built for a Cruise car to be $30 wholesale, $40 if you buy the exact same thing as an aftermarket replacement part, and $150 if you buy a standalone LIDAR sensor with the parts you need to actually use it.
Because a replacement Nexus 7 screen isn't going to natively interface with my Raspberry Pi, and neither will a Cruise LIDAR sensor. A $30 part that needs a $20 interface isn't any better than a $50 part that includes the interface.
If autonomous vehicles are a success, then there may be surprisingly few totaled cars in the future, compared to what we see today. I'm excited for that and agree with you that some really cool stuff can come out of the salvage of the future.
It's interesting to think about the situations where even the most sophisticated sensors and software may be unable prevent an accident that can total the vehicle. Animal crossings, catastrophic vehicle part failure, surprise road debris, etc...
If you want to play with this stuff today, comparatively simpler 2D scanning LIDAR can be found cheaply on Ebay.
You can usually find used SICK units there for well under $500.00 USD (I once scored one unit for $250.00); even when they look in terrible condition they usually function fine (they are designed for industrial environments and usage). These old units have a downside in that they use older serial comms (RS-232 and RS-422) that can be tricky to get going; for one unit I had, I had to send it back to SICK to get it reset to RS-232; it was originally set to RS-422 (faster), but to go back to the slower standard required a converter, which I didn't want to spend the money on if the unit didn't work properly. Then - on top of all that - I had to set up a Window NT 4.0 workstation to use the software. Plus they require a 24 volt DC power supply (great for mobility chair robots, though!). Once I got it running, though - everything worked well.
They are also large - as in coffee pot sized (in fact, they are kinda "affectionately" termed this), and made of aluminum casting and other metals that make them relatively heavy. This also makes them very robust, and are perfect for medium and larger-scale robots (no desktop rovers here - these sensors work well on robots the size of 24 VDC mobility chair bases and larger); for instance, if you look up Stanley from the Darpa Grand Challenge 2005 competition, you'll see five of the units facing forward on top of the vehicle. Each scans a line out many meters in front of the sensor, with a field-of-view (FOV) of about 180 degrees. It's like a radar sweep. As the platform vehicle moves forward, if the sensor is angled down, you can build up a height-map of the terrain in front of the platform. It is also possible to mount the sensor on a turntable to rotate it while it is oriented vertically, and scan a 3D volume, though at a much slower rate.
The other low-cost possibility is to purchase a surplus Neato robot vacuum LIDAR sensor. These units are also 2D, but use a parallax shift method of determining distance, kinda like explained here:
Instead of a webcam, a 2D line-element imager is used. The whole contraption is then mounted on a turntable with a slip-ring system (for signals/power) to scan a 360 degree view around the sensor - the whitepaper behind the sensor can be found:
These sensors - as pulls or replacements - can be found on Ebay for about $100.00 USD each, and are much smaller (desktop rover sized), but have a limitation in that they don't work well outdoors (sunlight swamps the sensor) - but given the device they were meant to be used on, an indoor vacuum robot, that wasn't a concern.
Another option people have tried has been to get a low-cost 4-6 conductor slip-ring off Ebay (cheap), then build a similar motor-driven scanning system like the Neato lidar uses. For the actual sensor, a "LIDAR-Lite" module is used, which are relatively inexpensive (more than a Neato module, less than a SICK). They also have a longer range, and work well outdoors in sunlight. These have been mostly "homebrew" systems, so googling for such projects will be the best way to find them, but I've given enough info here that you probably won't need much hand-holding to implement something similar.
A recent off-the-shelf option that has become available, and is still fairly low-cost, is the Scanse Sweep:
At a price of $350.00 USD, it isn't as cheap as a Neato sensor, but it does work better than that unit while still being small with longer distance capability, and is much cheaper than a new SICK LIDAR unit.
Finally, a form of LIDAR that is somewhat like a combination of computer vision coupled with lasers, this project has always been of interest to me:
It is similar to the old webcam LIDAR I pointed out earlier, but uses a laser to create a line (cylindrical lens element), which by displacement viewed in the camera, and some simple trig, can lead to a form 2D and 3D scene reconstruction.
The original unit described used a standard NTSC camera and a basic detection circuit, but today one could potentially use a high-resolution web camera, and a laser pointer with a line diffraction grating, plus OpenCV to do the detection work. It is similar in principle to the simpler 3D turntable scanners out there used for 3D printing and model capture, if you need another example.
So - I guess all I can say is that if you are interested in this stuff, there are many ways to play with it today, without having to wait for future junkyard finds. While I agree that such sensors will be neat to play with, and will only come down further in price as they become more ubiquitous for manufacturers, you don't have to wait...
Your last example is a standard laser triangulation rig. They're extremely simple to build and give excellent accuracy. Here's an example system I was involved with for cast steel: https://doi.org/10.1109/CRV.2016.55
The pedant in me wants to point out that laser triangulation has nothing to do with LIDAR, even though it's commonly lumped in. It's basically a parallax measurement, whereas LIDAR refers explicitly to time of flight (either direct or by proxy from phase/frequency offset).
Mostly they're used for profiling on conveyor-based systems - anything where the target moves underneath the stripe. The downside is usually a big trade off with accuracy and depth of field; and usually their usable range is poor (> 5m and you're struggling and have eye safety problems). You can buy nm-precision systems, but they only work within a range of a few cm. Also the triangulation angle is important - the further away you go, the wider your laser-camera baseline needs to be to retain accuracy.
https://spectrum.ieee.org/cars-that-think/transportation/sel...