Let me start this post with a brief history of autonomous driving. Most people date the start of the autonomous driving era to 2004, with the Grand Challenge in the Mojave Desert over a 150-mile course. That turned out to be 142 miles longer than necessary, since the furthest any vehicle got was eight miles and the $1M prize was not awarded. Perhaps a more auspicious year to pick as the start of the autonomous driving era would be 2005, when the second Grand Challenge took place, with the prize money up to $2M. Five vehicles finished the entire 132-mile course. (I wrote about this in one of my first blog posts here Ten Years Ago Self-Driving Cars Couldn't Go Ten Miles.) The third Grand Challenge a couple of years later involved all the vehicles, along with other vehicles with professional drivers, all driving around a disused US Air Force base at the same time. Many vehicles successfully negotiated the course, obeying traffic laws and avoiding other vehicles. In just a few years, the technology had advanced so fast that everyone assumed we'd all be driving autonomous cars within another ten years...by 2017 or so.

Huge sums of venture capital were poured into startups, all the individuals on the teams from CMU and Stanford were hired, and all the big car companies (OEMs in the jargon) along with the biggest tier-1s kicked off their own programs. With a normal automotive development cycle being five to seven years, there was suddenly a feeling that it was already too late. Google (now Waymo) self-driving cars became a common sight in Mountain View. Uber self-driving vehicles in Pittsburgh. Cruise around San Francisco.

In a post about the Automobil Elektronik Kongress in Ludwigsburg in 2017, I wrote Automotive Software Development Used to End with SOP where the executives of the automobile companies attending were largely focused on the cultural change necessary to transform their companies to ones which could undertake agile software development, take on some of their own electronics, and more. In 2018, there was the first fatal accident, which I wrote about in In Other News, 100 People Were Killed by Cars Driven by People. It was becoming clear that getting the basic technology to work was a lot easier than getting the final few percent to make it completely safe in all circumstances, or at the very least ten times safer than human drivers.

By 2019, at Ludwigsburg, companies were becoming more circumspect about their investment, focused on a shorter timeframe with less automation, both to mature the technology, and gradually acclimatize the public to the idea of autonomy. I covered this in a post Ludwigsburg: It's All About Return-on-Investment, and one from the Automotive Summit, too, AImotive: Shifting Gear in Automotive.

But things continue to advance. The picture at the start of this post is a robotaxi in Las Vegas during DAC, my first experience of being in a fully autonomous vehicle, albeit with a safety driver (and, unfortunately, a ban on photography inside the vehicle).

Arm

Later that day Robert Day, Arm's director of automotive solutions and platforms presented The Next Big Step in Autonomy: From Prototype to Production. Arm divides the progress up into three areas:

• ADAD/autonomous, which involves sensor processing for lidar, radar, and camera, along with in-cabin driver montoring
• Digital cabin, which is enhanced instrumentation, navigation and infotainment, and passenger multi-screen systems
• Connectivity, which is LTE and 5G (cellular), V2x, over-the-air software updates, and high-precision GNSS (GPS)

He mused on whether full autonomy is going to be evolution or revolution, since current mobility platforms (robotaxis) are very different from current ADAS platforms. He thinks that the move to full autonomy will likely involve lessons from both, and cautioned that:

the leap from level 3 features to level 4/5 is orders of magnitude in complexity

This wasn't just his and Arm's positions. He had several quotes that I think capture the current autonomous zeitgeist:

Waymo CEO John Krafcik: "Autonomy always will have some constraints.”
Ford CEO Jim Hackett : "We overestimated the arrival of autonomous vehicles.”
GM executive: "The engineering challenge of our generation.”
May Mobility CEO Edwin Olson: "Robotaxis are likely to be a fantasy until 2035”

A report from Forrester from which Robert showed a slide (above) showed the challenges in reducing SWaP-C while maintaining performance. SWaP-C (highlighted in yellow on the slide above) is code for keeping the size of the system (S), the weight of the system (W), the power low (P), and the cost down (C). The odd punctuation comes about because it used to be SWaP (originated in the military) for size, weight, and power, and then the cost aspect was added. Automotive is a lot more cost-sensitive than mil-aero, who will pay a lot to get SWap down, especially anything that has to be put in space or something that flies. Anyway, you can see that those four challenges are among the most difficult in Forrester's list.

Or course Arm is there to help. I assume you know Arm's basic product line (although if you ever work out the numbering rules, then let me know). For automotive, they have one innovation that they call Split-Lock. This allows two cores to be locked for safety, where they run the same code and each operation is checked against the other core. For performance as opposed to safety, they can be split and run different code. When Arm presented this technology at the Linley Processor Conference, they were asked just how dynamic this is. Basically, the cores need to be booted up either locked or split, it is not something that can be turned on and off (mainly, I suspect, because you'd never be able to get the operating systems back in sync to lock once they had been running split). The first two processors with this capability are the Cortex-A76AE and the Cortex-A65AE, which is also Arm's first multi-threaded processor. One key thing about the Split-Lock capability is that it is transparent to the software.

One problem with two processors is Segal's Law:

A man with a watch knows what time it is. A man with two watches is never sure.

Of course, this is a joke, but it does have an important truth in it, which is what to do when the Lock breaks. The safety systems in many aerospace applications are triply redundant with majority voting to avoid this problem. In automotive, there is expected to be some sort of safety processor that can take over when uncorrectable failures are detected. Functional safety (FuSa) standards like ISO 26262 require some means of recovery from safety-critical failures.

Of course, it wouldn't be Arm without a broad ecosystem of partners and open-source support:

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.