Recode has obtained detailed internal documents showing that Uber has a long way to go to create working self-driving cars. Uber’s vehicles, according to the documents, still need extremely frequent human intervention, lagging far behind competing systems.
The most negative number for Uber is that, during the week ending March 8th, human drivers had to take over from autonomous systems once every 0.8 miles. That number has actually worsened slightly since the end of January, when drivers had to intervene every 0.9 miles. Those numbers are for all such “disengagements” of driverless systems, including a car misreading lane markings— a seemingly minor problem that nonetheless was behind a recent Tesla crash.
Get Data Sheet, Fortune’s technology newsletter.
In the most recent period, human drivers had to take over only once every 200 miles to prevent a serious accident, a number which has improved significantly. This category would include incidents like missed stop lights in San Francisco, which the company later mischaracterized in public statements.
Despite slight improvement, those rates compare poorly to similar metrics from other self-driving car projects. Waymo, Google's self-driving spinoff, reported only 0.2 disengagements per 1,000 miles driven in California in 2016. Cruise, a self-driving startup now owned by GM, reported 18.5 total disengagements per 1,000 miles driven.
One mitigating factor for Uber’s disengagement numbers is that many of those miles were logged in urban Pittsburgh, arguably a more challenging environment than, for instance, a California highway.
The other takeaway from Recode’s report is in the number of miles logged by Uber’s self-driving systems. That number has risen steadily as Uber’s self-driving test fleet has grown, from 5,000 total miles per week in early February, to just over 20,000 miles per week in early March.
Those numbers compare slightly better with the competition. Waymo has put at least 636,000 miles under its autonomous vehicles’ wheels in California, while GM’s Cruise only logged 9,776 miles in 2016.
But the real 500-pound gorilla in this room is Tesla, whose vehicles have have logged 222 million miles in Autopilot mode, and gathered more than 1.3 billion miles worth of driving data. That’s not quite the same as actual autonomous miles driven (Autopilot is, despite its name, not full autonomy), but the data is a huge asset for machine-learning purposes.
Of course, Uber is working from behind in the race towards autonomy. Its Advanced Technology Center in Pittsburgh, staffed by poached Carnegie Mellon researchers, only launched two years ago. Tesla’s Autopilot was first offered in October of 2014, while Google has been working on the problem since 2009.
But that, arguably, is precisely the point. Uber can’t afford to remain years behind the competition on what it has described as an “existential” necessity. Tesla, Waymo, or GM are certain to deploy self-driving taxi fleets the second their tech is viable, or license it to someone who will. In the low-friction ridesharing marketplace, that could cut Uber off at the knees.