Tier One Supplier Valeo has a new LIDAR and put it on this delivery vehicle being made in China Brad Templeton While CES is officially the Consume
While CES is officially the Consumer Electronic Show, LIDAR and other components for robocars have become a major part of what is displayed there. There were at least 43 companies exhibiting LIDAR products at CES, and some sources claim there are now close to 150 different companies out there in the field. Among the companies, there is surprisingly little duplication of design — almost every one takes a somewhat different design, and each believes their approach has a chance at being one of the few winners.
The main news this CES was the “arrival” of a robocar LIDAR from Bosch, and higher performance, lower price LIDARs from many companies, plus a few new design entrants.
LIDAR companies are looking to win in one or more of the following spaces:
- The brass-ring is to be the primary LIDAR for a full robocar.
- Others seek to be a LIDAR for driver-assist “Autopilot” systems, though some of those (like Tesla) don’t use a LIDAR.
- Shorter range and cheaper devices can provide “close in” views for low speed navigation and seeing things close to a vehicle the main LIDAR doesn’t.
- Some companies hope for good business in other areas like low-speed robotics, smart city sensing and security applications.
Particularly in the first two categories, there are many important methods of differentiation, all aiming at the following goals:
- Cost: Taxi fleets can afford high-cost sensors, but private cars can’t. And everybody wants to pay less if they can.
- Range: Range of 200m or more is necessary for highway. 1550nm LIDARs get this easily, 905nm and similar need to work for it. Of particular interest are the dark, hard targets like tires on the asphalt.
- Robustness: Everybody wants to be sure the device keeps working and stays calibrated in the rough, vibrating automotive environment. Many early LIDARs did not.
- Resolution: Finer detail, particularly in specified areas of interest, like the horizon, or a particular obstacle.
- Field of view: Some LIDARS scan a full 360 degrees. Some have a narrow field which means you may need several of them. Some use a narrow-field long range LIDAR to focus only on the road ahead, and use something less for everything else.
- Speed of operation: People want at least 10 scans/second, or even 20-30 if they can get it. There is usually a trade-off between frame rate and resolution or field of view or cost.
- Special features, such as measuring the speed of targets, avoiding interference, and doing a “flash” to capture the entire view at once can be a plus.
- Size: Being smaller and easier to integrate into the vehicle design is a plus, though early adopters may actually prefer more obvious sensors like roof-mounted 360 degree units.
- Decoding: Both better hardware for decoding the return signal, and being tied with special software to help interpret the point clouds.
To attain all these goals, LIDARs differ on what wavelength they use, what type of emitters and detectors they have, what optics they put in front of those, how they steer their beams (if they steer them at all) and how they process the signals.
The laser wavelength makes a big difference. Operating at 1550nm (long infrared) allows much more power to be used safely, for much better range. It’s also expensive because you can’t use cheap silicon approaches. In short-infrared, the shorter the wave the more efficient silicon is, but the more ambient light there is from the sun. It’s hard in these bands to get the fully desired range on the hard dark targets, though many vendors claim they pull it off. Everybody can see things like retroreflectors (as found on signs and some lane markers) but black cars, clothing and tires are another story.
Beam steering is another big area of differentiation. Several LIDARs, particularly the oldest and 360 degree designs, just spin the whole LIDAR in a circle. You will also see many devices with a small mirror that is able to vibrate (often in two dimensions) to steer the beams. Another popular approach is to use very tiny “MEMS” mirrors which can be built on a chip. (This is often called solid state but it has small moving parts.)
The real solid state approaches include “phased array” steering (common in radar) and frequency based steering (using a laser whose frequency can be quickly changed and a prism, which makes light change direction based on its frequency.) New entrant Bajara uses this method for steering one direction. Strobe, which was acquired by Cruise, is rumored to also use this method.
“Flash” LIDAR does not scan. Rather it has a very large number of sensors, and possibly emitters to do the whole scene at once. Traditionally that’s been expensive and also hard to get the range from since the flash involves immense power. Thanks to cheap VCSEL laser arrays on silicon, new companies like Sense Photonics hope to win with this approach, though they currently can’t deliver the range for highway driving.
These solid state approaches are sought after because it is believed they will be the most reliable and robust in the harsh automotive environment. Large moving parts are harder to keep robust and to keep calibrated. Nonetheless, as you might guess, every vendor now insists their current products are robust and won’t need frequent replacement or servicing — and automakers will demand that. Robotaxis that come home to base every night can tolerate less robust instruments though only if that offers some other great benefit.
The most talked about new entrant was Bosch. While Bosch teased they would launch their new LIDAR at CES, they pulled back and revealed no details, other than it would be “long range” and be at a price point suitable for the driver assist market. (Generally, the driver assist market needs sensors that are well below $1,000 since you don’t want to add many thousands of dollars to the cost of a vehicle people buy off the lot. The taxi market can tolerate much more expensive sensors, spreading that cost over many customers at a few cents a mile.)
People pay attention to Bosch because it is one of the world’s premier top level automotive suppliers. Nobody has better reach in selling to car OEMs, who all know the giant company well and trust it. That’s a big leg up on small and unknown startups. If Bosch has a decent LIDAR it may well beat better LIDARs from small companies.
According to Bosch staff, they examined every LIDAR provider they could find out in the market, hoping to find one which they either could acquire or work with as a partner. They found all the approaches out there lacking in some way, and elected to build in-house. They claim that their design is not identical to any other design out there — a remarkable claim considering how many different designs are there. Their claim that they did not find anybody to acquire may stem from the very high valuations that companies in the robocar space have commanded.
We’ll have to wait to see more about what Bosch is actually building.
Aiming at Driver Assist
Another theme to the show was the “Robocar Winter” pull-back several companies expressed by now aiming at the driver-assist market. While most early excitement was about LIDARs for full robocars and robotaxis, most companies now feel early volume will come from selling cheaper, simpler LIDARs aimed at doing competitors to Tesla Autopilot (which does not use LIDAR, but which has had several crashes which would have been prevented by LIDAR.)
It is true that auto OEMs will sell many driver-assist systems before they sell an end-user real actual working full self driving car to an end-user. The latter is very hard, since end users won’t like cars that only function in certain areas, and they don’t bring the car back to the shop every day for a chance to improve it, the way robotaxis would operate. A product like Tesla’s autopilot has become a must-get feature for higher-end cars, who are all being outsold by Tesla. LIDAR can make such a product safer, faster if it fits the price point. Not every OEM has the willingness to do such a product (or more advanced standby supervised products that let you not pay attention to the road on the highway) without a LIDAR.
Almost all companies are now forecasting they will get sub-$1,000 LIDARS out when they are selling in quantity. A common number cited by the 1550nm crowd was $500 to $1,000. Some of the 905nm crowd spoke of goals in the $200-$300 range. Chinese makers promote the low cost of Chinese manufacturing as getting them to these prices first. One of the more impressive players was Livox, which was quoting their LIDARS from $600 to $1200 in single quantity today.
While it is true that everybody needs to promise these low prices, enough companies seem confident of it that one can now be more sure of the prediction that low cost LIDAR will be available from one vendor or another within a couple of years, in time for most robocar forecasts. This is important for the “cameras vs. LIDAR” debate, where one of the primary arguments for camera approaches has been the high cost of LIDAR. Camera-only players like Tesla are betting that they can get computer vision to work well enough. LIDAR players are betting that LIDAR will be cheap. The latter bet seems very likely to pan out. (The debate is not as simple as this, but cost is an important component. People like Elon Musk believe LIDAR is a distraction at any price, saying that “LIDAR is a crutch.” At the same time, computer vision currently has only one leg.)
Returning the speed
A handful of companies are offering to provide the speed of any target seen by the LIDAR. This is usually done with the Doppler effect, which can be calculated when the device uses “Frequency Modulated Continuous Wave” the same way most automotive radars do. Last year, an FMCW company named Blackmore was acquired by Aurora. Several other independent companies, such as Aeva and a few others are selling the same approach.
It’s quite useful to know how fast a target is going. With traditional LIDAR you need to look at several LIDAR frames to figure out speed. This can mean 100 to 200ms of latency for the frames, plus extra processing time of 100-200ms. That can make the difference in crucial circumstances like a surprise obstacle on the road.
Luminar, a leading player in the 1550nm space, has developed an alternative approach. They are sending multiple laser pulses at the objects of interest just a few milliseconds apart. That’s enough to tell how fast they are moving if you are accurate enough. This (or FMCW) means you might identify a stopped object in a few ms with no computation instead of taking 400ms as it would with vision or ordinary LIDAR. Some recent dangerous Tesla accidents involving impacts with vehicles parked incurring into the left lane show the importance of this determination in some cases.
Each company has a story of why they will win. Here are the factors that will determine the winner:
- 1550nm’s longer range and power will prove superior to shorter-wave devices, but it comes at a significant cost.
- Simpler devices with no moving parts, or smaller moving parts (such as MEMS) will be more likely to be robust.
- Devices which can mount the lasers, emitters and optics on one chip or extremely firmly are more likely to be robust
- Companies with history and experience selling to Auto OEMs, like Bosch, and the few LIDAR players with existing large-unit orders from OEMs, will have a leg up in the channel. Robotaxi products will be less affected here.
- Those who can manufacture at scale, with good quality, at lower prices, will win in the long run. Anybody who can’t do this won’t be able to play at all outside of small volume Robotaxi no matter how good their instrument.
- In the robotaxi business (which comes first) more established 360 degree designs may dominate, if they become robust, because they mean having to buy one or two such units.
- Special abilities like returning the speed of targets may become crucial for full robocars. People will pay quite a bit extra if it avoids important classes of accidents.
There will be more to say about LIDAR in the coming months. This includes an effort to build a database of all the major LIDAR providers and more examination of the big LIDAR+computer vision vs. computer vision alone battle.