Radar measurement, characterization, and testing of next-gen 140-GHz systems in Imec’s DSP/RF laboratory in Belgium. (Imec)

Making sense of next-gen ADAS sensing

Experts at AutoSens 2023 explain how technology is right-setting the industry’s driver-assist and autonomy ambitions.

“I don’t expect [SAE] Level 3-capable vehicles to ship in significant numbers until 2030,” asserted Rudy Burger, of Woodside Capital Partners, during May’s AutoSens 2023 conference in Detroit. Burger, one of the automated-vehicle industry’s most respected analysts, keynoted a blunt assessment of an AV market that continues to rebound in this “unusual and uncertain time” post-COVID.

The industry’s focus this decade clearly is on enhanced ADAS (what some dub ‘SAE Level 2-Plus’) for the majority of passenger and light-commercial vehicle applications. Nearly half (46.5%) of all new light vehicles currently sold in the U.S. are L2-equipped. Level 3 and 4 autonomy development is aimed mainly at robotaxi, local goods delivery and over-the-road haulage operations — a smaller but still- robust space, said Burger and other experts who spoke with SAE Media at the conference. Cost, immature technologies and the sheer number of new players entering the automotive sensor market have conspired to cast off dozens of startups.

“Our database tracked over 120 lidar ‘companies’ funded over the last five years,” Burger noted. “More companies were created than the market could absorb” — a byproduct of what he calls “ADAS Sensor Fatigue Syndrome.” From a high of 37 ADAS sensor companies launched with funding in 2016, the total dropped to 16 companies in 2021. As of early May, no new players had yet to appear in 2023. “It’s tough to raise money now,” he said. Notably, six of the top 17 publicly funded ADAS tech companies are Chinese.

More TOPS, please
As OEMs map their product strategies beyond the current product cycle, M&A activity in the ADAS/AV sensor and compute markets remains vibrant. Key incumbent Tier 1s (notably Bosch, Continental, Gentex, Koito, and ZF) are among the most active investors. Suppliers interviewed at AutoSens project the typical L3 consumer vehicle will feature 8 to10 cameras, five radars, one lidar, ultrasonic and moisture sensors and a compute capability of 100x TOPS (trillions of operations per second, a primary measure of the maximum achievable throughput).

For the much smaller fleet of L4 robotaxis, each be fitted with up to 14 cameras, 21 radars (10 short range, 8 long range and three articulating types) and five lidars; L4 is said to require ten times the compute capability. Power-hungry lidars with 1550-nm wavelengths now in development will need 1000x TOPS, for example. 

More TOPS typically means more silicon area, more power and more cost. “Software will be the dominant focus in automotive in the next five to 10 years,” Burger stated. But the auto industry still is in the “dark ages” on software development compared with the PC industry, he added. As it moves to the “software-defined vehicle,” a more universal end-to-end operating system is needed.

“I think Tesla will continue to lead the way because they are vertically integrated – they own almost all of the tech in the stack” — there is no comprehensive off-the-shelf O/S solution, Burger said. 

Cameras are evolving so that optical depth will come for free, experts said. In radars, engineers are working to overcome various signal limitations, including the challenging “under bridges” environment that often causes false braking. Evolving 4D imaging systems are expected to dominate the ADAS market, according to Abdullah Zaidi, leader of Rivian’s sensors team. He said 4D radars deliver 300 m (984 ft) of detection range, with 80-deg. angle of resolution, versus current systems with 18-deg./200 m (656-ft) range. More advanced chipsets are bringing higher resolution; Texas Instruments’ latest offers 1192 virtual channels while Arbe’s offers 2304, Zaidi noted. Gated imaging and short-wave imaging radar (SWIR) are soon to arrive in volume.

SiPM and SWIR
Lidar will play a significant role in the passenger-vehicle segment, with new edge-emitting and fiber types now in key suppliers’ pipelines. “We see long-range lidars being adopted sooner than short-range types,” Zaidi, formerly with Qualcomm, told the AutoSens audience. Work is ongoing to reduce lidar power consumption to below the 12-13 Watts of the most efficient types, he said.

Experts agree that lidar’s main holdup is cost: To see it in volumes greater than 200 million units annually, the technology will need a significantly greater amount of light power — and that’s expensive, they noted. On cost and performance, the hardware is evolving. Regarding detectors, for example, the industry is moving to silicon photomultiplier (SiPM) technology, Zaidi said; SiPM is a solid-state, high-gain radiation detector that produces an output current pulse upon absorption of a photon. These sensors with single-photon sensitivity can detect light wavelengths from near-ultraviolet (UV) to near-infrared (IR).

He explained that SiPM has evolved into a premier photodetector, noted for its high internal gain and low excess noise. Further, SiPMs are mechanically compact, can be operated at low voltages and are impervious to magnetic fields. The energy resolution possible with SiPMs is in some instances comparable to that achieved by photomultiplier tubes (PMTs).

Semiconductors, expected to account for 20% of a premium vehicle’s bill of material by 2030, according to Roland Berger, play an increasingly key role in creating more-capable and cost-effective sensors. Belgium-based semiconductor developer Imec recently introduced a miniaturized, single-chip, solid-state lidar solution operating in the SWIR range, noted Steven Latre, the company’s head of AI research. 

“SWIR is close to natural light,” he explained. Imec also is developing SWIR imaging using QD (quantum dot) sensors for vehicle occupant sensing. Latre sees a trend away from sensor-data fusion toward what he calls “sensor-tech fusion,” with camera, radar and lidar evolving into a single unit, but admits that discrete sensors give OEMs flexibility in locating them on the vehicle. Echoing other industry experts at AutoSens, he stressed that compute requirements increase “dramatically” when ADAS capability moves from L2 to L4.

“Higher levels of [vehicle] automation will require more TOPS. The amount of compute that we need is growing exponentially,” he said, to support more advanced sensing across different modalities.

Latre believes automotive “semis” will evolve as part of the current E/E architecture revolution, from general purpose single-die SoCs to more domain-specific solutions. Imec is moving to “chiplets” — smaller chips with lower power requirements based on a SoC architecture that can reduce the form factor of lidar modules, making them more package-efficient (Imec also is targeting the aerial drone and medical markets). The company is developing lidar chiplets for wavelengths around 1550 nm, range resolution of a few centimeters and capable of performing imaging at several frames per second, at 10-30 Hz. Faster time to market is another benefit of chiplets, Latre said.

Mercedes’ L3 pioneer
Konstantin Fichtner’s much-anticipated presentation on Mercedes-Benz’s new Drive Pilot system attracted a large audience at the AutoSens conference. Touted as the company’s first SAE Level 3 (currently for Germany only) production vehicle (and among the industry’s first), Drive Pilot builds on Mercedes’ incumbent driver-assistance L2 package by adding lidar, an advanced forward-looking stereo multipurpose camera (SMPC), a mono camera facing rearward, a microphone array for detecting emergency-vehicle lights and sirens, and a moisture sensor — all equipment for the S-Class, Maybach and EQS vehicles equipped with Drive Pilot. A new antenna array that provides more-accurate vehicle positioning is included.

Nevada is slated to be the first U.S. state to allow L3 vehicle operation (via state regulation Chapter 482A for Autonomous Vehicles). According to Fichtner, the Drive Pilot L3 project lead engineer, operations will be expanded to California later in 2023, with the company’s “ambition” to offer an L4 system by “end of this decade.”

 Fichtner defines L3 as “conditionally automated driving.” He admitted that driving an L3 system, “you get to love traffic jams” for the relaxation afforded when the system takes control. Now including more than 16 driver-assist subsystems, the S-Class likely is the world’s most complex ADAS-equipped production vehicle.

“With every drive in a Drive Pilot vehicle, 60,000 signals are recorded” through 10 data loggers, Fichtner said. As other AutoSens experts noted, onboard compute power requirements will rise with the ingestion and processing of more data. “We might need active cooling” in the future, he said.

Continue reading »
X