Does The Tesla Crash Show That Current ADAS Liability Models Retard Innovation

In “Tesla
TSLA In Taiwan Crashes Directly Into Overturned Truck, Ignores Pedestrian, With Autopilot On,”  Brad Templeton from Forbes discusses the recent Tesla highway accident. In his excellent analysis, Brad asks all the important questions relative to the accident. Why didn’t radar collision avoidance work?  Would LIDAR have helped? What about the camera system? He goes through a well reasoned argument about the potential causes and technical solutions, and one of his major conclusions is “… missing a giant truck and a pedestrian is a bit much for a system which is also the foundation of a purported “full self driving” system.” It is very difficult to disagree with his conclusions.

Building even just a “simple” robust conflict avoidance system is a non-trivial task. ADAS designers face the following difficult challenges: 

1. RADAR:  In the airplane context, RADAR is very effective because every object encountered is of high interest. However, in the ground context, RADAR signals reflect off the environment in such a manner to create blind-spots and “ghost objects.”  
2. CAMERA: It is easy enough to get a picture, but interpreting it is quite challenging. Today’s machine learning algorithms are the equivalent of google image search, so if the object has not been seen and recognized, the ADAS system does not know what to do with it. 
3. TRACKING:  Both radar and camera are fixed on the car, which makes them susceptible to situations of road curvature, hills and valleys from a field-of-view point-of-view.
4. DRIVER INTENT:  Is the driver accelerating towards a car with the intent of a passing maneuver or is the driver distracted ?   How do we decide ?

With some understanding of the problem, one has to have sympathy for the ADAS Designer’s challenges.  The response to these recognized challenges has been a permissive regulatory environment whose objective is to encourage innovation.  For ADAS,  this permissiveness is provided by the statement that the driver is ultimately responsible.  However, is this “crutch” actually retarding innovation?

Let’s consider the behaviour it seems to engender: 

1. Vague and Unclear Functionality:  What exactly can the customer count on when buying an ADAS system?
2. Lazy Validation:  Partial validation and partial coverage are the norm in the industry today. By the way, how exactly do we know that the next revision of the software is actually better than the last one from a safety point-of-view ?  

Both of the above are covered by the “loophole” that the driver carries the ultimate liability. Further, the current structure creates a predisposition against braking. Why?  Premature braking clearly attaches liability to the ADAS system while a more permissive (aggressive) approach just shifts liability to the driver. There are upcoming technical solutions to many of the issues mentioned.  These include:

1. Focus: “What Can Tesla And Waymo Learn From A Human’s Ability To Focus” discusses the notion of abstraction and focus. In this world, obstructions are not recognized by a google image search, but as higher abstract objects. In this view, one does not have to recognize the big object on the road is an overturned truck, but just know, it is a big object.
2. Communication of Risk:  “Is Progress In Autonomous Technology Gated By Research In Animal Communication?”  discusses the importance of communication in the driving task. One of the key ideas which must be communicated is the notion of risk. If an ADAS system expects to have a human take over on-demand, it seems reasonable to expect communication from the ADAS system of its perceived level of confusion. 
3. Driver Expectation: “How Safe Is Safe For An AV ? The Answer (Expectation And Communication)” discusses the importance of setting clear expectations of behaviour in very easy to understand terms. Expectations are the foundations of safety. How can the driver access an issue without clear expectations of the ADAS system’s expected behaviour?

Indeed, all of the above led to a strong argument for measurable safety outlined in  “Measurable Safety, The Missing Ingredient To Demonstrating ADAS Value.”  This article argues that measurable safety would show customer value and build a monetization path for ADAS.  More rigour around safety metrics could be driven by a regulator or by an industry consortium such as AVSC. However, in the context of today’s driver liablity “crutch,” there is little motivation to drive in this direction, and innovation itself may be retarded in the process.