Tesla's new Full Self Driving beta 9.0 still looks a bit too exciting

Tesla has started releasing its much-hyped "Full Self-Driving" beta 9.0 software to selected early-access users, upgrading their Autopilot systems with new off-highway autonomous capabilities and better visualizations.

As recently as 2004, self-driving cars were the stuff of science fiction. The first DARPA Grand Challenge robo-car race was widely reported as a bit of a farce, when no contestant got further than 7.1 miles into the 142-mile journey. Just one year later, five teams finished the race.

Peak autonomous vehicle hype kicked in sometime around 2014, when Tesla first switched on its fledgling Autopilot system, which was able to self-steer and maintain a safe distance on the highway. But while human learner drivers tend to view high-speed highway driving as a frightening milestone to conquer, for autonomous systems, the relatively controlled environment of a highway is low-hanging fruit. The real challenge is dealing with the chaos of city and backstreet driving, where anything can happen at any time, and frequently does, with potential threats coming from all directions.

According to Gartner's famous Hype Cycle, autonomous vehicles have plunged in recent years from the "peak of inflated expectations" into the murky depths of the "trough of disillusionment" when the vast majority of laypeople stop being excited about a new technology and only the die-hard nerds continue to give a fat rat's backside about it. This, of course, is typically when the really hard work starts to get done.

And that's what Tesla has been grappling with lately. While many manufacturers have contented themselves pulling sheets off non-functional, driverless robo-pod concepts at car shows, Tesla has controversially pressed ahead with its strategy of getting things out into customer cars as soon as possible, holding firm to the mantra that if it's safer by the numbers for the car to drive than a human, that's what should be happening.

So far, it has indeed been safer. Teslas have now driven more than a billion miles on Autopilot, and the data paints a very rosy picture that continues to improve every quarter. From the company's 2021 vehicle safety report: "In the 1st quarter, we registered one accident for every 4.19 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.05 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 978 thousand miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 484,000 miles."

Sprinkle a grain of salt on these figures, as Autopilot is designed to hand control back over to the driver when things start to get hairy, but the point remains: it's doing a pretty darn decent job for a six-year-old technology.

On Saturday night, the company began the slow release of the latest major upgrade to its Full Self Driving (FSD) system. FSD v9.0 is the first of Tesla's software versions that ditches radar to run a "Pure Vision" system relying mainly on camera vision. Radars, said Elon Musk in a Q1 2021 earnings call with analysts, are "one of the last crutches" for autonomy, and they're occasionally responsible for annoying and potentially dangerous glitches.

Version 9 has only shipped to about 2,000 cars at this point, members of the exclusive early access beta testing list. And this is still very much a beta release; the release notes are clear that it "must be used with additional caution. It may do the wrong thing at the worst time."

The system includes a new "mind of car" visualization system to show what the car's seeing as it wends its way through the streets. This system is basically there to give drivers a way to quickly visually assess whether the car's seen what they have, and in the new release, objects like cars, trucks, brake lights, traffic cones, lane markers, pedestrians, dogs and cyclists are rendered in greater detail, with brightness and translucency used as visual cues to show how "confident" the pure vision system is that it's identified and categorized things correctly. It sure looks neat.

Version 9, according to the release notes, can now "make lane changes off highway, select forks to follow your navigation route, navigate around other vehicles and objects, and make left and right turns." Tesla cars will also now watch drivers through cameras above the rear view mirror to make sure they've got their eyes on the road – but the company says these images "do not leave the vehicle itself, which means the system cannot save or transmit information unless you enable data sharing."

"Difference between v8 & v9 is gigantic," tweeted Elon Musk back in May – and that may be the case under the hood, but for most drivers thus far this has been more of an incremental improvement. Early access testers are reporting a general improvement in the system's ability to navigate and drive its way around town, that it seems to be correcting itself quicker when it's about to make a mistake, that it's creeping forward just like a human driver when it can't quite see around a corner, that it's handling complex situations like multi-lane roundabouts surprisingly well, that it's squeezing its way around tight, blind corners better, that it's doing a great job in foggy conditions, and that it's taking appropriate caution around lots of pedestrians.

On the other hand, some situations still give it the willies – like turning left across traffic onto a six-lane road without traffic lights. That's understandable, most human drivers aren't a fan of that one either. It sometimes waits too far back from an intersection, and testers seem to need to give it the odd prod with a blip of the accelerator to let it know it's waiting for ghosts. Here, certainly, erring on the side of safety is clearly the right way to go.

In tighter streets, it's acting very nervous around pedestrians walking close to the street, as you can see in the video below – but it's also no pushover, refusing to let people in willy-nilly when it has right of way.

Others are finding more significant issues, including one guy who's finding the system consistently wants to drive him right across the path of high-speed traffic, and who is convinced he'd have been t-boned half a dozen times already if he'd let the car do what it wanted.

So there's absolutely no confusing v9 with a fully finished level 5 autonomous car; anyone running this build still needs to stay very awake. "Use Full Self-Driving in limited beta only if you pay constant attention to the road," read the release notes, "and be prepared to act immediately, especially around blind corners, crossing intersections, and in narrow driving situations"

There's no word yet on when v9 will be rolled out to a wider number of US drivers, and certainly no word on when it'll be ready in other countries. But Elon Musk has stated that his goal is for the system to be ready for level 5 (hands-off, eyes-off, feet up on the dash, take a nap) by the end of the year.

Of course, Elon Musk is known to be just a teensy bit optimistic when it comes to Tesla targets. Tesla's own director of Autopilot software CJ Moore isn't convinced, and nor is the DMV. Interestingly, Moore seems to have some lofty targets before he'd consider it ready: 1 to 2 million miles of driving between human interventions. That does not feel close watching these early beta testers.

It seems more likely that the age of the human driver is ending not with a bang, but with a drawn-out series of whimpers. Such is the nature of beta-testing extremely complex software in the real world with genuine safety implications. Personally, I'm staggered by how fast Tesla is dragging this technology out of that "trough of disillusionment" and into the "slope of englightnment."

This stuff is still glitchy enough to be scary and exciting, but it's starting to feel like it won't be long before it's completely mundane and boring. And once it's mundane and boring, it'll be ready to change the world just as fast as road authorities will allow it to. Just a few billion more miles' worth of machine learning, edge cases and human correction to go...

Source: Twitter