Affiliate Disclosure
If you buy through our links, we may get a commission. Read our ethics policy.

Apple greatly reduced autonomous vehicle testing in 2019, fleet of 23 cars drove only 7,500 miles

Apple's self-driving car testbed in California.

Last updated

Apple's autonomous vehicle testing program saw a significant drawdown in 2019, with its fleet driving 72,201 miles less than it did in 2018. The company also appears to have changed its disengagement reporting protocol, information commonly used to gauge system efficacy.

According to numbers submitted by Apple to California's Department of Motor Vehicles, 23 of the company's 69 registered self-driving testbeds were active between December 2018 and November 2019. Collectively, the fleet of modified Lexus RX450h SUVs logged 7,544 miles during the reporting period, down from 79,754 miles in 2018.

Apple reported 64 disengagements across the 7,544 miles driven, which equates to 8.48 disengagements per 1,000 miles. This compares to a whopping 69,510 disengagements, or 871.65 disengagements per 1,000 miles, recorded in 2018. At the time, Apple's system was reported, on a disengagements-by-mile basis, as the worst-performing test platform in California. By contrast, then-front runner Waymo managed 0.09 disengagements per 1,000 miles over the same evaluation period.

While the improvement seems impressive, it should be noted that disengagement reporting is not a hard science. Typically, disengagements are defined as instances in which a self-driving vehicle's autonomous systems are unable to process current conditions, forcing it to pass control back to the human driver. Human interventions also count as disengagements.

As explained by Apple in a letter to California's DMV last year, the company in 2018 took a "conservative" approach to disengagements that required "drivers to proactively take manual control of the vehicle any time the system encounters a scenario beyond our currently proven abilities." Further, the company said its "software self-monitors and returns control back to the driver when it encounters any errors or issues during operation."

The letter implies Apple's built-in threshold for failure is, or at least was, much less tolerant than competing systems, a thesis was borne out in total disengagement figures recorded during the period. From December 2017 through November 2018, the 47 other firms testing automated vehicles on California roads clocked a cumulative 3,040 disengagements over nearly 2 million logged miles.

That said, Apple appears to have narrowed its reporting methodology in 2019.

Of the 64 disengagements, 52 were initiated by the autonomous vehicle system, while human pilots overrode AV decisions 12 times. The most common reason for disengagement was a vague "controls discrepancy" issue.

According to numbers compiled by CNET, Baidu led the pack in 2019 with 0.06 disengagements per 1,000 miles, followed by Waymo and Cruise, both of which managed 0.08 disengagements per 1,000 miles. Toyota reported the highest rate of disengagements with 1,620 per 1,000 miles, though the Japanese carmaker fielded six cars that traveled 1,817 miles.

Apple's self-driving car initiative began under the "Project Titan" banner, with the program initially tasked with creating a branded car from whole cloth. Apple scaled back operations in late 2016 after hitting a number of snags and the Titan team has since refocused attention to autonomous vehicle subsystems. Whether an "Apple Car" will one day make a public debut is unknown.



18 Comments

mr lizard 354 comments · 15 Years

At the time, Apple's system was the worst-performing test platform in California.”. 


Based on what evidence? Human intervention counts as a disengagement. It’s entirely possible that Apple’s testers are directed to intervene more often out of an abundance of caution. I doubt they want to be all over the news for being involved in a collision, unlike some of the other folks. 

EsquireCats 1268 comments · 8 Years

Yeah it's hard to compare disengagements across brands, but you can certainly get insight from inside each project to see how quickly they're resolving perceived issues.

It's also difficult to read into reduced road time, it could mean anything, such as time spent in R&D, time spent testing new software on private tracks before moving into public spaces.

Since Apple now have roads in continental USA 3D mapped as part of their Maps overhaul, it seems probable that this data can also be used in some virtual testing scenarios for the car project.

avon b7 8046 comments · 20 Years

mr lizard said:
“At the time, Apple's system was the worst-performing test platform in California.”. 
Based on what evidence? Human intervention counts as a disengagement. It’s entirely possible that Apple’s testers are directed to intervene more often out of an abundance of caution. I doubt they want to be all over the news for being involved in a collision, unlike some of the other folks. 

Most of what you have mentioned, was covered in the article.

The evidence is there but the caveats and changes since then are also explained in the article.

Until two different cars are put to exactly the same road conditions at the same time, it will be difficult to truly compare how autonomous they really are and how they perform in other areas. 

GeorgeBMac 11421 comments · 8 Years

I don't see how a Post-Jobs Apple could hope to compete against Musk.  

The success of both was largely / mostly due to an overwhelming and obsessive drive to accomplish a lofty goal.   Musk has it.   Steve did too.

larryjw 1036 comments · 9 Years

No one should read much of anything into this article. While the article mentions three variables  — disengagements, miles driven, and company — there’s likely hundreds of variables that would go into understanding the status of self-driving vehicle systems. 

Those in AI suggest general such systems are more than ten years out. Until then, we will need to be content with bots vacuuming our carpets, delivering pizzas to college dormitories on research campuses.