Uber declares itself devoted to safety


This weekend, Uber released a long and detailed "safety report" with some of their learnings and new plans after their fatal error. I have not had time to read it all yet due to travel, and will offer more detailed comments later. You can read the comments of the Washington post and many others online.

Even without reading it, I am sure the document will be contrite and declare a dedication to safety. And I am even confident that Uber will become much more dedicated to safety than before, since they have had more of a wake-up call that any other team.

They go into more details on the technical flaws they want to improve, including the latency in their perception system and of course the need for better emergency braking ability.

But even without reading it, I want to reiterate a point from before. Uber's failure came from two very different components, and both were bad but only one was the true failure. That was the failure to have good safety driving -- both in terms of Uber's protocol of managing and training safety drivers, and the performance of the particular safety driver they hired.

While Uber's technical systems performed particularly badly, because not seeing a pedestrian alone on an empty road is an unacceptable level of performance, all prototype vehicles regularly encounter issues where the intervention of a safety driver is needed to prevent an accident. All of them, particularly in their early years. Even the best ones, like Waymo. They would not be prototypes if this were not the case. Of course, they can differ in how often this happens and how severe the incidents are, but they will happen.

The main public safety concern revolves around whether the safety driver protocol and team of drivers is good enough to do those interventions when they are needed. Here, Uber failed badly. This includes failures in their protocol, but also includes technology failures, in that they failed to build system to alert an inattentive safety driver when needed, such as when emergency braking is needed, or if attention is insufficient. Uber outlines plans for some of this in their plan to return to the roads To get back on the road, Uber does not need to show they have a perfect car, because nobody does. They have to show they have a very good safety driving system and team. A system which understand that humans are not perfect either, and is ready for their mistakes as well as mistakes of the system.


I don’t understand why the internal emergency braking systems (like Volvo one) are disabled?
It’s always better to have a reserve system, and also these systems are design to act only in emergencies - that means they’re shouldn’t have too much false-alarms.
And they are used with human drivers, who are unpredictable in accident conditions, why don’t use them with a more predictable robots then?

Generally, any robocar team is building a vastly superior system than any existing ADAS tool. You don't want two systems trying to control the car. You could consider that for the brakes, but you still could end up in a situation where the ADAS system hits the brakes for a false positive, and the main system doesn't want to and can't "un-hit" them. It would have to plan for that.

The right way to do it is that Uber should have an independent AEB which is neither the Volvo system nor their main system. The independent system would be simpler than their main system, but use the vastly better sensors available to it, unlike the Volvo ADAS. Another good approach would be to have the AEB signal from the Volvo ADAS be an input to the main system, or the independent system. In that case it would be, "The ADAS AEB is telling me to brake. Unless I have a deliberate reason not to brake, I should do so."

But generally, a well designed robocar will already have a better version of the AEB and so using the existing one just adds complexity. The Uber was not such.

Add new comment