Velodyne Plans a Lidar Megafactory
spectrum.ieee.org
spectrum.ieee.org
Velodyne is still doing this with rotating machinery. The more advanced concepts are true solid state devices. Continental, the auto parts maker, bought Advanced Scientific Concepts' expensive, but successful technology (Space-X uses it to dock with the ISS) and is trying to make it cheaper. If they can get the cost down, that's probably the best technology around.
Quantergy announced a no-moving-parts LIDAR last year, but is having trouble delivering. (Despite that, their site says "Quanergy is the leading provider of solid state LiDAR sensors." No, they're not, not until they ship the product.) Quantergy recently announced a mechanical scanner, which is apparently shipping, and now claims that their solid state unit will ship in volume in 2017. We'll see if they deliver.
Somebody is going to get this working soon, now that there's a market. I first saw a flash LIDAR as a demo on an optical bench at ASC back in 2012. That was clearly the right approach, but it was a long way off commercially.
[1] http://epc-co.com/epc/EventsandNews/News/ArtMID/1627/Article...
A drawback of flash lidar is that the light spread out with an inverse square law. With diffuse reflection, the returned power decays in an inverse zenzizenzic fashion [1]. Lidars with a collimated laser beam are far less affected by the inverse square law on the emission side, therefore getting better range and signal-to-noise ratio.
[0] https://en.wikipedia.org/wiki/Matched_filter [1] https://en.wiktionary.org/wiki/zenzizenzic
Our new systems are palm sized, and will get much smaller shortly.
What happens if you have a few dozen Lidar-equipped cars in the same spot? Is there something like crosstalk, scatter or noise between the cars which result in Lidar artifacts?
[1] http://ideas.4brad.com/tricking-lidars-and-robocars [2] http://www.aiqus.com/forum/questions/22373/using-multiple-li...
If one did want to think adversarial one could explore a system receiving Lidar, say from a car behind, figuring out its transmission and doing your own illumination to confuse or jam the other car system. But the chances for this state arising naturally are astronomically small. My 2c.
Every material that is black could be a good candidate. I am wondering if the google cars do have a problem with cars that have a matte black finish (those PVC foils).
But you have to keep in mind that they are not invisible in the real sense, because you get a hole in your scan where you don't have points in your pointcloud. This could be detected, because the Velodyne usually has many beams pointing downward which eventually would hit the road, but would leave a hole instead if there is an NIR absorbing material.
And if you use sensor fusion, you can use a camera to check for dark colors in those areas that might help you to detect such cases
That's why you want to profile the road from a high LIDAR. It's not just about sensing obstacles. It's about building a height map of the environment. This was more important in the DARPA Grand Challenge, where we had to drive off-road. Potholes and rocks were to be expected. There are visual illusions for both vision and LIDAR, but few of them present the appearance of a flat road.
But I am not saying that this will become a huge issue, because it would affect individual measurements, and the overall scan should be fine, just a bit noisier. And noise is something they have to deal with today as well due to different reflectance properties, electromagnetic interference, flying voxels, etc.
If someone really wants to cause an accident, it's kind of hard to stop them, regardless of whether you're in a self-driving car or not.
Reflected sunlight caused us a number of problems but I'm going to assume that any lidar used on a car is rated for outdoor use. I suppose those might still give bad readings if you're driving directly into the sunset or something, though.
1: https://techcrunch.com/2016/11/15/new-lidar-package-makes-it...
https://www.microsoftstore.com/store/msusa/en_US/pdp/Kinect-...
https://www.asus.com/ca-en/3D-Sensor/Xtion_PRO_LIVE/
(launching soon:)
http://click.intel.com/intel-realsense-developer-kit-r200.ht...
there could be phones that include these sensors
If you have a great business plan that needs 3D consumer-priced 3D scanning, I'd start with a Kinect if I were you :)
Nonetheless, progress appears to be made in the area. The article states that "Google originally had to pay US $73,000 for each unit, itself a rotating array of 64 laser beams". Today a hobbyist can just buy an 1D LIDAR for 150-200$ or a 2D LIDAR for 450$ on seeedstudio.
If yer not picky, you can generally pick up a cheap "repair part" or "surplus" Neato 2D Lidar unit from ebay, for about $100.00 or so. They are generally pulled from broken Neato vacuum robots. Alternatively, a used Neato (or sometimes even a new one) can generally be had for under $450.00 (plus you get a ton of other robot experimenting parts!).
Another option for hobbyists are used SICK serial-port lidars; I've picked those up for around $300.00 off ebay. They run on 12-24 volts (great for vehicles and electric wheelchair-based platforms), and use RS-232 or RS-422 for communications (where it becomes a PITA is when they are set to the RS-422 protocol - because then you need a special adaptor - and not all of the adaptors work - to reprogram it - using a piece of software that runs on Windows NT4).
They are very robust units, though. Big, and heavy (not for small robots - not at all; these units are about the size of a coffee maker). There are three main types - each with a different color case (yellow, blue, and grey). One of them is an "area coverage" only sensor (and doesn't have a way to output scan data), while the other two are more generalized (at least, this is what I recall - I probably have some details mixed up).
These are on the used market because companies are phasing them out, I believe, for the newer ethernet based models.
Posits that we don't need each and every auto manufacturer to be building and developing these systems (sensors + AI). Just a handful or two of large players (and they need not necessarily be auto manufacturers themselves) all building and developing to minimum national standards.
Basically like the driver theory test these pieces of software must all pass the same criteria and standards. They can do better, they can set the bar higher for themselves, but there needs to be an pretty exhaustive check-list of compliance.
That way when you're buying a car you're not wondering which edge-case is going to get you or someone else killed.
There are going to be fatalities and injuries along the way, that's for sure, but so long as the number of fatalities and injuries is well below the current horrific rate this tech will be beyond revolutionary.
There's been plenty of demonstrations by various players that self-driving vehicles can get a great amount if not nearly all needed information from camera-based systems, and drive well using them.
Other sensors will be needed, of course - and maybe LIDAR can even play a part there; the more data you have, from more sources, the better - this has always been true in robotics.
I just think that LIDAR won't play a primary role, but rather as a backup or accessory system component.
But that's precisely the mechanism that's been working and accepted since the invention of the directed machine...
(Also, just to clarify: a stereoscopic distance determination isn't meaningfully a "guess". That's been a solved problem for a long time.)
The goal can't ever be perfect safety. The machines themselves aren't perfectly safe even if they aren't driven into anything. It just has to be measurably safer than a human driver.
And in any case , sensor fusion seems to be a common approach, so we'll probably need both lida and cameras.
so, ban humans from driving?
I'm really surprised that one of them hasn't done it already. Hokuyo especially - one of their sensors is already relatively small, and USB based; I wish I could get my hands on a used one (so far, all I've seen on the market are older SICK units).
It's probably because they are all entrenched in the niche industrial market; they may not be able to make a lower priced sensor without cutting into that market (but if they don't, they might see that market erode anyhow).