Don't Miss
Home / Blog / Uncertainties and Automated Driving

Uncertainties and Automated Driving

Technological advancements in the face of computational power, sensor technologies, more robust algorithm techniques and fields like machine learning and deep learning are the major enabling force of fully automated driving.

Automation has been a salient achievement of research in the last century and will rightly be an important aspect for the industry in the coming decades. The last two industrial revolutions (IR3 and IR4) can be attributed to automation in systems . IEEE names Industrial Revolution 4.0  as “cyber physical systems” and includes big data, autonomous robots, autonomous vehicles and system integration, among others. The trend towards automated driving has increased and is projected to increase further in coming years. Level of autonomy is also expected to enhance from driver assistance to fully automated driving.
Technological advancements in the face of computational power, sensor technologies, more robust algorithm techniques and fields like machine learning and deep learning are the major enabling force of fully automated driving. However, modeling system and specifically real time systems is always ambiguous; one has to conciliate what has been modelled and what is left deliberately or even unknowingly. As Albert Einstein said:

“As far as the laws of mathematics refer to reality, they are not certain; and as far as they are certain, they do not refer to reality.”

Though, it might be arguably true that he was not talking about autonomous vehicles (or may be he was, who knows), its applicability on the autonomous vehicles case cannot be denied. Autonomous cars, as they have been envisioned, will operate in a highly uncertain world. Moreover, as many of the components are represented by probabilistic models, their reference to reality (perhaps Einstein was talking about autonomous cars after all) cannot be fully justified. In short, uncertain systems are envisioned to be deployed in uncertain world, a vision shown in Fig. 1.

Figure 1 Dream of Urban Mobility. Picture courtesy: Bosch Group & Daimler.

For many readers at this point, it is a very alarming situation. But is it really so? Fortunately, it is not as bad as it seems. We live in an uncertain world and as it is said “there is nothing certain but uncertain” and so far we have managed well with the inclusion of new technologies in our lives. Autonomous cars are a special case though; they are safety critical. What if a certain amount of light in the sensors of the cars blinds them or they create blind spots owing to the narrow vertical range. The results can be catastrophic as was the case of Uber . One natural way to deal with this problem is to reduce those uncertainties to a minimum where their effects can also be minimized to an acceptable level.

Figure 2 Uber Volvo XC90 after the accident 

Uncertainties can be categorized in many ways, however the most common being the division in aleatory uncertainty and epistemic uncertainty.
Aleatory uncertainty is the inherent randomness of the process. In the words of great theoretical physicist, cosmologist, and author Stephen Hawking it is the phenomena of “God plays dice”. This uncertainty is irreducible (how can one reduce the phenomena of “God playing dice”). It is represented using probabilistic models.

Figure 3 Aleatory Uncertainty: Not only does God definitely play dice, but He sometimes confuses us by throwing them where they can't be seen. 

Epistemic uncertainty is the lack of knowledge that is represented in the models of the system. This uncertainty is considered reducible. Nassim Nicholas Taleb in his book “The Black Swan” uses the term “known unknowns” and “unknown unknowns” which can be attributed to epistemic uncertainty: The state of we know that we do not know and we don’t know that we do not know. The “unknown unknowns” of Taleb, are also categorized as Ontological Uncertainty in some literatures.

For the sake of writing a cliché literature, let’s take an example that represents both aleatory and epistemic uncertainty. Let’s make it more clichéd by taking a scenario of an autonomous car situational awareness, which is being provided by a camera sensor. The car is designed for urban deployment. At a given instant, what actors (the participants of the environment) will the camera see? Since the movement of actors in the open world (urban environment in which our vehicle is) is random (remember, God likes playing dice), we cannot answer this question with complete knowledge, however, previous data can help us model a distribution.

Suppose the camera is modelled to recognize some actors and leave some (“the known unknowns”). If need be, the left actors can be modelled. However, on the other hand if some actors were not known at the deployment time they can only be discovered after the deployment. These actors can then be assessed based on their rate of occurrence and need of modeling, and then either modelled or left.

In order to argue that epistemic uncertainty has been minimized to an acceptable level, models depicting the same are required. “What is the acceptable level” is beyond the scope of this blog.  Different modeling techniques are available to model uncertainties based of Bayesian theorem, evidence theory or subjective logic, however, there are imminent challenges that needs to be addressed.

Autonomous cars provide a way forward for automotive industry. They not only provide lucrative economic benefits but also envisage the dream of safer, connected and energy efficient transport. Uncertainty quantification and minimization of such system is important as it effects the safety directly. If right efforts are not made in this regard, safety can become the tailback of the autonomous cars revolution.


About the author: Ahmad Adee

Ahmad Adee received his M.Sc. in Advanced Robotics from Politechnika Warszawska (Poland) and Ecole Centrale de Nantes (France) in 2018 as part of European Master in Advanced Robotics (EMARO). He is the Early Stage Researcher in the project MSCA ETN-SAS (Safer Autonomous Systems). His objectives are to explore propagation paths of functional insufficiencies in software intensive systems using Model Based System Analysis (MBSA) techniques, with the inclusion of probabilistic analysis. The research activity is conducted in Corporate Sector Research and Advance Engineering, Robert Bosch Campus, Renningen.