Uber meaninglessly announces they are back on the road, when they aren't


Uber made a strange announcement this week, that they are back on the road in Pittsburgh, when they actually aren't.

The announcement is a strange PR stunt. It's full of the list of things they have changed since their fatality. They are the things everybody expected: Two safety drivers, driver attention monitoring, better safety driver training and quality. Except there's one catch -- none of this matters because they are just driving the cars manually, for mapping and scenario gathering. Uber has made a big announcement that they have people driving cars manually around town -- which of course Uber has a million of around the world, without special procedures. You don't need safety drivers or training to do that.

The list also mentions that the automatic emergency braking (the built in one that comes with Volvos) is enabled. They have always had that enabled when doing manual driving, and I will presume that when they some day resume autonomous testing, it will be disabled again.

What this announcement really represents is a way to pretend to taking an intermediate step towards the thing everybody is scared of -- Uber resuming full autonomous testing with their bad track record. They are testing the waters with a non-announcement.

When they do get ready to do the real deal, these details will make more sense. We will also want to know about what they have done with their classifier that did not figure out what she was until very late, and the obstacle trajectory predictor which failed to figure out the victim was crossing the road and on an intersection course. We'll also want to know how many false positives their perception system is generating, and whether it is still so many that braking at >0.6G is not enabled.

No report on whether the system will now give some sort of alert or minor brake jab when a possible-positive is visible, to make sure the safety driver is paying attention. (There is a risk to this if the safety driver starts imagining they now can look away, but the gaze detection system should forbid that.)

Uber might have also considered something more radical -- three safety drivers. After all, there is no reason the software operator has to be in the front seat. You could have a 2nd safety driver in the front passenger seat only to be a 2nd set of eyes on the road. You could give her a 2nd set of control (like an airplane copilot) but I don't think that's needed. The software operator does need some view of the road to compare the scene with the view of what the car is seeing, but that can be done almost as well in the back.

Nobody else does that. For everybody else, the safety record has been excellent, and there is no call for it, and it's a bunch of added cost and a bit more difficulty for the software operator, so I don't think anybody else should. Uber, though...

I also hope Uber has done a bunch more testing of pedestrian scenarios using balloon pedestrians on their test track, as well as in sim. They need to get their false positive number down to a reasonable number.

In other news

Much better news for Waymo, that has announced 8 million miles of on-road testing and hopes to get into deliveries, logistics and public transit. They just put special pick-up spots into Walmart parking lots in Phoenix, where cars will take customers to Walmart. This may a little cheat of its own -- Waymo has said it will offer self-drive service to the general public this year, and "ride to Walmart" would count -- but not be particularly exciting. Let's hope that's not what this is.

Ford has upped their plans to put an extra $4B into robocar projects, and Cruise has taken money from Softbank and may be spun out to give extra value to GM. Mercedes is talking about splitting itself in 3 (with one "moibility" unit) which shows the automakers are finally starting to realize how dire their position is.


One takeaway from the Uber fatality is just how financially expensive and damaging to reputation such a tragedy can be. No easy way to measure it but long term the fatality has likely cost them billions. So it possibly makes sense for Waymo to start slowly in the safest areas (geo-fenced areas in outer suburbs) within the safest cities (Good climate and well laid out road network). Shopping areas and outer railway stations with inadequate parking could be good places to start. The latter could encourage more people to use the train service and ease congestion further towards city centres; no harm in reducing congestion in order to win some goodwill in the early stages.

Everybody would like to start with the easy problems first, and move upward. Well, almost everybody -- Cruise has declared a goal to start in the dense and complex streets of San Francisco, and claim it as part of their edge in the game. And some overseas teams are also making the case that rising to the harder problem is the way to get to commercial success faster.

Waymo is so far ahead that every team has to declare something they are doing different in order to not seem like a follower.

I have to admire Waymo, not only are they trying to compliment public transport by proving a solution to the 'last mile' problem, they are also planning on helping those with disabilities. They seem well ahead of Uber in their approach to building goodwill with city authorities and trying to win over an an often skeptical public. Many of the criticisms of autonomous vehicles are that they will extend a very city degrading car culture. It would certainly help the politics if some environmentalists saw autonomous vehicles in a positive light rather than something to be opposed. The 'cars are bad and rail is good' theme runs deep and I hope Waymo can show this thinking needs to be challenged.


The automatic emergency braking is on when the car is being driven manually? I thought the problem with it was false positives, hitting the brakes at random times (which seems like a good way to get rear ended). How is this any better when the car is being driven manually?

The Volvo that Uber uses has "AEB" which is an ADAS (driver assist) feature, as a standard function, as well as several other ADAS functions from Volvo and their supplier. Many cars have this. ADAS automatic emergency braking is not fully reliable, because it is always used with a driver. Generally, if testing a robocar, you turn other systems off as you don't want competing systems trying to control the car.

A robocar itself inherently has an emergency braking system (automatic is redundant here, it's all automation.) You would not even call it AEB, because it's not a distinct function, it is central to the system that if it detects an obstacle that must be avoided, it will attempt to avoid it, mostly by stopping.

Uber's system, however, had a false positive problem -- braking too much for ghosts. Regular ADAS systems avoid the false positive problem by accepting more false negatives. That's why a Tesla autopilot (which is also ADAS) will happily plow into the broadside of a truck or a road barrier.

With a robocar, that's not acceptable. But in prototype development, Uber decided that while it would not accept false negatives deliberately, it could not allow its system to apply hard brakes over 0.6g, and depended on the safety driver to do that.

Add new comment