The promise made by the U.S. that driving fully-automatic by 2020 would not be a dream of a far away future, did not come to fruition. But nevertheless, Germany is well on the way to become the first country worldwide permitting automatic driving (Level 4) legally within predetermined areas of operation of public transport. In the centre of all this: The Berlin start-up Motor AI that has specialised in observing and maintaining the safety- and legal requirements as specified by the European Union. Motor AI is developing the intelligence for autonomous vehicles of Level 4 and 5. Contrary to machine-learning systems based on trained situations, the Motor AI system is to make safer and more traceable decisions during completely new scenarios. #ki_Berlin spoke to co-founder Adam Bahlke.
Hello Mr. Bahlke. You founded your AI-start-up Motor AI together with Roy Uhlmann. What exactly does your product stand for and how did you personally get involved with autonomous driving?
Motor AI developed a system for autonomous driving – the Autonomous Driver – consisting of perception, sensor fusion and decision making. Our USP and our original idea is ‘Decision Making’.
Since 2017 we have been working at the decision making by means of AI-procedures and asked the question, how a system can make decisions without having to carry out training of all sorts of possible decision scenarios with immense amounts of data. Simply speaking, we asked the question, why a person does not have to go through billions of kilometres to practice for his driving test, but is able – after 20 to 30 practical driving sessions – to solve new problems in traffic reproducibly time and again.
Originally, all we wanted to do was to solve the so-called “edge cases” in autonomous driving. But we realised very quickly that in doing this, we held a decision framework for the overall “Decision Stack“ in autonomous driving in our hands. With regards to the development process it seemed logical to us to develop also perception- and sensor fusion, leading us to our current product, the “Autonomous Driver” for the ÖPNV [public transport] and logistics.
With your product you count on Cognitive AI instead of Machine Learning. How do you explain this and what are the advantages of your approach?
To start with, cognitive AI is a process instead of a specific model. Let us look at Deep Learning first [Editor’s note: a method of machine learning]; it is presumed that there does not exist an a-priori meaning of the data that go into a model. Therefore, various mathematical processes are carried out to create a meaning. Sometimes it works very well, sometimes one obtains a model that is of no use – but in respect of both cases we humans do not know why or how the model reached this decision. And it is at this point that cognitive AI starts with the contrary principles. The data have an a-priori meaning and they have to be processed in certain logical procedures. And although deep-learning approaches have an advantage in respect of perceiving the surroundings, the cognitive AI has a huge advantage regarding decision making, where it is mandatory from a legal point of view to act rationally, meaning in a way that can be explained and is causative.
The advantages are clear to see, are they not?
On the one hand we have made a decision for the cognitive AI as its way of functioning complies with German and European laws. And on the other hand, when making decisions in autonomous driving, there are so many potential scenarios that it needs a system that can recognise and solve these scenarios itself without having been trained for this. Compared to Decision Making, Perception only needs a relatively small quantity of data. An object with a data file of two million well annotated pictures is more than sufficiently trained. Deep Learning in this case is certainly practicable.
Based on the possible entities and variants, one assumes approximately one quattuordecillion scenarios in Decision Making, a figure 1 with 45 noughts. A deep-learning system has to be fed with an uncountable number of training data to solve sufficient decision scenarios – a truly Sisyphus task. And as long as they are not provided with sufficient training data, the deep-learning systems remain mere black box-systems in the autonomous driving based on probabilities.
You may recall that five years ago it was announced in California that autonomous vehicles of Level 4 would be driving on our roads by 2020. The reason why this did not happen lies in the endless variations of decision scenarios that were not considered at that time. A cognitive system overcomes this problem.
There are companies - above all in the U.S. and China – who rather prefer to work with a black box system, i.e. based on a non-determined algorithm. Will such an approach be certified by the TÜV [Technical monitoring association] authority?
Professor Wolfgang Wahlster – at the time CEO of DFKI – offered us the opportunity to be part of the DIN-commission for AI in the section of mobility. As a member of the commission we did not only rub shoulders all of a sudden with representatives of car manufacturers such as BMW and Volkswagen and TIER1, but also found ourselves sitting next to representatives of TÜV. This gave us an insight into the various perspectives concerning exactly this question.
There is basically a connection between those who had the task to formulate the laws for autonomous driving and above all consider the origin for such a decision and the technicians whose focus was on solving autonomous driving technically and who therefore watched the result of the decision. The technicians would, of course, have preferred that the laws were adapted to the technology. While making a decision regarding autonomous driving, one would prefer to focus on probabilities rather than on causal chains. A huge change considering that our legal principles have been based on causality since Roman times. The discrepancy between the two points of view lead to the “Trolley-Dilemma“ or the well-known ethical discussion.
In order to enable the technical verifiability again, one looks to compromise by making given functions certifiable. Motor AI, however, developed a software with the objective to obtain a fully causal verifiability in terms of a deterministic decision of the system meaning the finding of a decision which can be interpreted and reproduced.
Safety is always an important argument for an approval and the press media inform us time and again about accidents caused by autonomously driving cars, for example, in the U.S. Is it possible to guarantee higher safety standards with a logic-based approach of Explainable AI (XAI) such as yours?
For being accepted by society, safety is the decisive criteria. Nevertheless, there will be accidents with Autonomous Driving, we must bear that in mind. We therefore think that the German system, where an independent third party as technical inspector, applies the same standards all the time and to every system regarding Autonomous Driving, is of advantage for the competition globally. Whoever can offer a system that can be interpreted and reproduced, is ahead of the game worldwide, contrary to those systems that only comply with locally set legal frameworks and safety standards of their own making.
Let us talk about the legal framework conditions: The Federal Government – as first state worldwide - is planning to allow per law the fully automatic driving (Level 4) within predetermined areas of operation of public transport. The Bundestag has confirmed this act already and the resolution of the Bundesrat is also available. What does this mean regarding the Autonomous Driving in general?
It is expected that the act will be announced beginning of 2022 and will subsequently come into force. Following this, Autonomous Driving Level 4 will then be allowed for regular operation of the ÖPNV and logistics within the areas assigned by the administrative districts.
We believe that this step is groundbreaking. On the one hand the subject of Autonomous Driving will thus leave the testing mode and become a real economic factor. Theoretical, economic multiples of technology become real and the influence of the technology becomes measurable.
On the other hand this will accelerate legislature of other EU states. The German act is already considered a blue print for the EU, an economic area of the same importance as the U.S.
For this reason we expect venture capital to shift closer towards money, also towards European Autonomous Driving enterprises that are already achieving a turnover.
And what exactly does this mean for an enterprise such as Motor AI?
When we founded the enterprise in 2017, we were not at all sure whether a company such as ours would not be better placed in the U.S. Today we know that we are in the right place at the right time. We enable autonomous On-Demand-vehicles to work in regular operation for the ÖPNV in Germany by the end of 2022.
What applies to Germany may apply to the whole of Europe soon, is that correct?
That is correct. Over the years the bureaucratic and safety-conscious Germany has solved the questions about Autonomous Driving legally and with a European mindset. And it is in particular the close connections within the European automobile industry that will contribute to the fact that further countries will follow the EU relatively soon.
One tends to look across the Atlantic where Autonomous Driving is concerned. What is possible in Germany, where is work still needed? What about subsidies or testing areas for Autonomous Driving?
Where we are concerned, we always saw problems in lengthy applications for subsidies, coordinating with dozens of project partners, especially regarding the use of testing areas. We, as an enterprise, have always tried to work as self-sufficient as possible. Although there were always offers on the plate, we always hesitated to commit too much to a strategic partner. Cooperation: Yes. Strategic or even an exclusive partnership: No. Today, also by looking across the Atlantic, we are one of the few hardware-agnostic and independently working enterprises in the sector of Autonomous Driving worldwide – and by having done this, we created enormous advantages for us. Irrespective of whether we are talking about sensor technology or selecting the vehicle itself – we are free to choose components to be supplied and without being dependent on anybody. This is also one of our keys to success.
What does the Berlin eco-system offer you? Why is this location so attractive to you?
We did not move to Berlin to found a company and therefore we did not select it as a classic location. For us the surroundings are attractive, the existing start-up culture and its mindset. The size of Berlin is naturally attractive but the thus developing conflicting interests often prevent the introduction of innovations. The decision paths in Berlin are just far too long. We would like to see more flexibility.
The development of autonomously driven cars has accelerated enormously over the past few years and progressed a lot. Putting the sci-fi cliché to one side: What will traffic look like in 2030 or 2040 around the world?
Many colleagues in our sector have said that the change from a standard vehicle to an autonomous vehicle will be as revolutionary as the change had been from horse to vehicle in the first half of the 20th century. We think that this historic comparison is most suitable. It demonstrates how profound economic changes are going to happen at the interface of the autonomous vehicle with services that are provided and linked to the autonomous car in future.