The idea of a “consumer AV” in 2025 requires a leap of faith for most industry watchers...
Leading autonomous vehicle developers are keeping open the possibility that the mythical robotaxi could possibly become real in 2022. The idea of a “consumer AV” in 2025, however, requires a leap of faith for most industry watchers.
Count me among the skeptical.
A lecture by Mobileye CEO Amnon Shashua on Tuesday at CES 2021 was one of the more “believable presentations” ever given on the topic of AVs. This is the judgment of Egil Juliussen, a veteran automotive industry analyst. The family AV could be a stretch, and its launch might slip a year or two, noted Juliussen, “but it’s a good goal.”
During two CES speeches, Shashua acknowledged there are technologial and business obstacles that make the case for the AV a tough sell. The AV industry has yet to demonstrate how that business can scale.
The question is how AV developers could avoid the impression that their AV development efforts are mere “science projects.”
In Shashua’s opinion, Mobileye’s rivals lack a solid driving policy that could convince regulators. He also noted there exists no high-resolution map, covering diverse geographies, that “allows cars to interpret the world.” Also absent is the sort of true safety redundancy that could make AVs socially acceptable.
These three elements — which Shashua calls the “trinity of Mobileye” — have “separated us from the crowd,” he said.
Mean time between failures
Shashua bought some credibility for his claims by holding to a stricter concept of mean time between failures (MTBF). Juliussen pointed out, “Mobileye is talking about not just MTBF for hardware, but both hardware and software combined.”
Instead of miles driven or disengagements — measurements AV developers often use in their AV safety claims, Shashua compares the MTBF of vehicles driven by machines with those of human driven cars.
In an exclusive interview with EE Times last year, Shashua acknowledged, “It turns out that the human statistics [on driving safely] are quite difficult to beat.” When asked why he compares AV with human performance statistics, he bluntly noted, “Society could do without autonomous cars, but no society can do without human drivers.”
Looking at the divide between driving-assist and autonomous vehicles, Shashua noted during this presentation, “We don’t see it in terms of customer functions.” He explained that it is not about whether an AV can perform more sophisticated customer functions than a driving assist. For Mobileye, the deciding measure is MBTF.
Carmakers can have a system with a level 4, level 5 capability while keeping a driver behind the wheel. In that case, the car isn’t really L4 or L5. “It would be essentially a level 2 vehicle,” said Shashua.
But what happens when you pull the human?
Without a driver, “the meantime between failures should be astronomically higher,” noted Shashua.
This insight has driven Mobileye into looking at “redundancy” very differently. The goal of safety redundancy should be “achieving those astronomical levels of MTBF,” said Shashua.
Consequently, Mobileye developed two separate perception subsystems, one based on cameras only, another combining a lidar and radars (using no vision). Mobileye’s engineering team has made each subsystem “independently perform equally at a very high level.”
Shashua noted that Mobileye had to work a lot harder to achieve this, compared to a lower-level of fusion — combining cameras, radars and lidars into one — which most AV companies do today.
Robotaxi first, consumer AVs next
Mobileye is set to penetrate the robotaxi market in 2022. In the first phase of the robotaxi, Shashua said, “We have the best-of-class time-of-flight (ToF) lidars from Luminar in our vehicle, combined with 360-degree coverage by using stock radars. The combination of the lidar and radars has proven to be enough to create a sensing state with very high MTBF, he explained.
Mobileye, however, is aiming higher. “In 2025, we want to have our AVs reach the level of consumer AVs,” according to Shashua.
There are two reasons why Mobileye is insisting on consumer AVs. It would require Mobileye to develop solutions to two persistent problems of AVs. One is the high cost of AVs and the other is the limited turf where they can operate, otherwise known as well-defined Operational Design Domain (ODD). Unless the AV industry can make cheaper cars, it’s hard to imagine robocars taking off. Similarly, if a AV’s ODD can’t leave its own circumscribed neighborhood, why buy one at any price?
Shashua didn’t say so in so many words, but Mobileye is attacking both factors — cost and geography — that most AV companies are reluctant to discuss.
Mobileye will be using “stock radar” chips for its first robotaxi phase. But in order to create a three-way redundancy, Shashua wants each front-facing sensor (camera, lidar and radar) to function as a “standalone system.” In particular, he said, “We want more from the radar. Today’s radar doesn’t offer the resolution and dynamic range [it needs] in a complex setting.”
In short, Mobileye is banking on the radar revolution.
Given that lidars cost 10 times more than radars, vastly improved imaging radars able to function as a standalone system will become critical to sufficiently reduce the coast of AV.
As many industry observers note, the big problem with radars is that they are “noisy.” Aside from increasing the resolution and dynamic range, even a trickier part of radars is dealing with a “side lobe” that echoes around each physical target. “We need to separate those echoes and leave only the true targets” to increase the accuracy of the sensed data, explained Shashua. This is a critical step for the radar to produce “something useful, just like an image from a lidar or an image from a camera.”
The new radar Mobileye is developing is based on a chip with 2,304 virtual channels using 48×48 transceivers and receivers, which is no cakewalk. The challenge is that more transceivers and receivers mean a drastic increase in computational complexity.
Shashua noted that resorting to brute force, a radar chip would need to deal with about 100 terabytes of data. Mobileye, however, has developed “very advanced AI algorithms and approximations,” which enabled them to reduce this number to 11.
Second, what Mobileye has done is to work everything on the digital domain, allowing the team to “build filters that are much more accurate and powerful” than dealing in an analog domain.
Shashua said Mobileye is working with Intel on this. They are targeting 2024 or 2025 for production of the new radar chips. “We already have samples, but again, this is futuristic.”
Separately, Mobileye is developing its own FMCW lidar. It’s no surprise that Mobileye has focused on frequency modulation continuous wave (FMCW) lidar systems, Juliussen observed. Today’s ToF-based lidar companies might be eventually forced to leave the market, he noted, as doppler-based FMCW lidars start gaining share.
Aurora, a full AV stack company, acquired in 2019 a Montana-based FMCW lidar startup called Blackmore. The advantage of the doppler lidar is said to be its ability to detect both range and velocity.
Juliussen noted that FMCW lidars, capable of detecting objects in a longer rage, are particularly crucial for AVs — such as autonomous trucks — driving at high speed on the highway.
Last year during our exclusive interview, Waymo said it is getting ready with the development of FMCW-based coherent lidars as their second-generation home-grown lidar.
Others are also gunning for FMCW lidars, including Calif.-based SiLC Technologies. SiLC last year at CES demonstrated a proprietary silicon photonics lidar integrated into a head lamp by Varroc Lighting Systems.
Given the patents held by these FMCW lidar companies, “these lidars could get expensive,” speculated Juliussen. It makes sense for Mobileye to lean on its parent Intel, which has ample silicon photonics expertise — including its own silicon photonics fab in New Mexico, plus manufacturing expertise and IPs.
Mobileye’s Shashua, while calling FMCW lidar “the next frontier of lidars,” noted that it is a very difficult technology. The greater range of FMCW lidar is also a good news for improving the ODD of AVs by 2025, he added. Another advantage is the higher immunity to interference. Juliussen said that FMCW lidar’s advantage is its ability to provide its own unique signals coming back from objects. “There’s less concern for lidar signals interfering with each other,” which could occur with non-FMCW lidars.
Mobileye’s CES 2021 presentation revealed its meticulous plans and strategies for an advanced AV roadmap – complete with the company’s ambition to set its sights on the next generation of radars and lidars.
Unlike many startups who fail to walk the walk, Mobileye’s track record suggests it might be able to deliver on its aspirations. Before scoffing at Mobileye’s ambitious plan for consumer AVs in 2025, warned Juliussen, skeptics might listen to Shashua’s whole lecture.
I concur. My only caveat is that Mobileye might have tied the whole 2025 consumer AV happy ending too neatly with a bow. As Shashua alluded during his presentation, both chips — radars and lidars — “aren’t coming out tomorrow.” This stuff is still futuristic.