An interview with Emmanuel Chevrier, CEO of AVSimulation and Thomas Nguyen, Head of Automotive Domain of AVSimulation
Where are we with artificial intelligence in autonomous driving ?
E. Chevrier : an Autonomous Vehicle – like a human being – goes through 3 steps to make a driving decision. It first perceives the outside world with its sensors (LIDAR, Radars and cameras), it is the «sense» phase of perception, then it analyzes the information, it is the «think» phase of the reflection and finally it makes a decision, it is the «act» phase of the decision-making. One thing that to my knowledge has not changed since I became Chief Executive Officer of AVSimulation is that we do not use artificial intelligence to take control of the vehicle and for the “act” phase. We use predictive algorithms to make decisions in an autonomous vehicle. Intelligence
AI remains, to this day, a black box, used to train the sensors that are the means by which the vehicle positions itself in its environment.
Artificial intelligence will be used to improve image recognition. For example, this is the type of intelligence that Facebook and Apple use to recognize faces or text. This image recognition allows us to do supervised learning, with the aim of improving sensor detection, and not to take control of the vehicle or decide the driver’s destiny. It is therefore really on the “sensing” part that we will find artificial intelligence to improve detection software.
AVSimulation is not a car manufacturer, but a software publisher. To date, is artificial intelligence present in SCANeR?
E. Chevrier: it is, Indeed important to remember that our raison d’être is not to develop vehicles, but to help our customers to design, validate and certify autonomous vehicles, ADAS (Advanced Driver-Assistance Systems) and the mobility solutions of tomorrow. On the ADAS part, we will be able to train detection and analysis algorithms («sense» and «think» thanks to artificial intelligence with our synthesis images.
We are also looking at how, thanks to artificial intelligence, we could improve our road traffic software by using real-world traffic data to better simulate the latter. We can also imagine using AI to make our tools more intuitive, so that our field editor can “guess” what the user intends to do and can offer accelerators.
In SCANeR we find artificial intelligence on driver models that can be aggressive, cautious or normal. We have pilots who are able to drive vehicles in a certain way, independently, and even in difficult situations.
For example, a race driver who has to choose in the most optimal way a trajectory, the choice of his speeds… the pilot has an adaptation phase and begins to learn to drive his vehicle. He tries different types of speed, inputs, his behaviour is similar to that of a human being when familiarizing himself with a new vehicle. We have models based on artificial intelligence and fuzzy logic, but we can still make progress for even more realism.
Where does AVSimulation position itself in terms of offering on the subject in the automotive space ?
E. Chevrier: We are interested in a number of subjects concerning artificial intelligence. For example, there is a topic that deals with the use of SCANeR software to train sensors. To get there we need scenes that are very realistic. This is one of the reasons why we collaborate with Epic Games on the Unreal engine so that the images are photorealistic and thus indistinguishable from reality. We are also working with IRT SystemX to make 3D scenes even more immersive and to ensure that CGI images can be used to train sensors while avoiding poor learning due to the artificial nature of the images.
Road traffic and driving are different from one continent to another. Do you have research underway to be able to specialise or regionalise road traffic using artificial intelligence ?
T. Nguyen: Behind these issues, there are a number of techniques that lead to artificial intelligence. SCANeR’s road traffic model, which is there to simulate and model the behaviour of vehicles on the road, is based on an artificial intelligence technique called a multi-agent system Developed for more than 25 years and based on rules, this technique allows to give good results on the establishment of representative situations by simulating individual behaviors. In our case, this implies the choice of a speed, of a track, the fact of carrying out, or not, a passing. The multi-agent is the fact that each driver has its own logic and its own rules. Mixing several actors in the multi-agent allows us to arrive at a representative result. In a context of artificial intelligence this technique is very similar to Machine Learning.
The AD/ADAS Pack includes a set of functional sensors. Is it planned to include artificial intelligence to make them smarter or more representative ?pour rendre ces derniers plus intelligents ou plus fidèles ?
E. Chevrier: The goal would be to make them more representative. It is possible to study the use of artificial intelligence to add defects. Indeed, even if the user can introduce a fault dose, our functional sensors tend to be too “perfect”. This is not representative enough of reality as the real sensors all have a number of defects. One of our objectives is therefore to use artificial intelligence to learn the defects of real sensors and be able to introduce them into our functional sensors.
What differentiates the intelligence of AVSimulation from that of your competitors ?
T. Nguyen: As I said earlier, there is the multi-agent approach, with a notion of fuzzy logic for balanced decision-making. At each step of the process, there are several types of rules that are evaluated and weighted to ultimately make a decision. The advantage is that we managed to prove that this led to large-scale, individual, realistic and manageable behaviour.
We are not looking for pure randomness, we must be able to take control of the decision. Since the technique is based on rules, it is quite simple: you just have to take control of the rules or deactivate them.
Can you give me your perspective on the value and use of artificial intelligence in AVSimulation products ?
T. Nguyen: As part of the simulation, artificial intelligence makes it possible to converge more quickly on more representative scenarios and simulated situations and to bring a little more randomness. In traditional approaches we do not make artificial intelligence, we build models that often tend to behave always in the same way. Artificial intelligence brings more randomness and predictability, which can be very interesting for customers. The constraint is that the unpredictable must still be realistic and a minimum controlled.
E. Chevrier: It is true that it may seem surprising when we talk about autonomous vehicles to hear that artificial intelligence is not necessarily at the controls of the vehicle, but that it nourishes and actively participates in the creation of these.
In this regard, the leader of autonomous vehicles Waymo (company of the Alphabet group), is positioned around the intelligence that will drive the autonomous vehicle and explains that at them, the primary objective is to manufacture the best driver. This driver draws on all the experiences of the vehicles on the road. It also feeds on all the virtual kilometers that are traveled in their simulators.
One of the issues that needs to be taken into account is that AI is present even in smartphones and allows its users to have access to voice and text detection. No one was planning to use voice recognition, and yet today we often go through dictation to write our SMS.
Thanks to the Cloud, artificial intelligence is consolidated, centralized and all learning benefits others. This is exactly what Waymo does, which will, thanks to the network and deep learning, have the most experienced driver in the world. This allows us to see that the collection of virtual and real data feeds algorithms that will make better decisions through AI.
At AVSimulation, we simply need to understand artificial intelligence as a tool that will allow us to make better traffic, photorealism, automate boring tasks. With the emergence of new standards like ISO 21448 it will be necessary to guarantee the security of the functionality autonomy (Safety of the Intended Function) in all circumstances, those that are known as the one that are not. By creating complex and realistic coherent virtual worlds thanks to artificial intelligence we will be able to generate more and more situations and decrease the unknown share. With artificial intelligence becoming more democratic, I think we need to look at what we can do for our customers and what data we have the right to use.
Waymo and others have announced that the autonomous vehicle will take longer than expected. Can you confirm that ?
T. Nguyen: Indeed, there is a large number of so-called “autonomous” vehicles. However, if at first there was a fantasy around the fully autonomous vehicle likely to co-exist with non-autonomous vehicles, able to pick up a person on their own and then transport them from one point to another etc. Engineers and legislators were quickly confronted with the harsh reality. However, there are more and more vehicles that are increasingly autonomous in certain controlled conditions.
This is the case, for example, in traffic jams, where the Traffic Jam Assist can take the hand, when, stuck in a traffic jam we must remain vigilant without really being able to do anything. So yes the fully autonomous vehicle will take time to arrive. That said, AI and simulation will allow for the smooth introduction of increasingly autonomous vehicles, which will, among other things, educate drivers and ensure a smooth transition to new forms of mobility.
T. Nguyen: This is an area where artificial intelligence could be used in a relevant way. Explore must indeed answer the question of the combinatorial explosion. In each scenario there are a large number of input parameters that can all vary independently from each other. If we had to run all the scenarios that could be generated, we simply couldn’t do it because it would take too long (even in simulated time)
The big challenge is to be able to determine what a relevant scenario is and determine the most effective method to reduce it. Another question for our R&D team is: How can we avoid deleting scenarios that might have been useful ?
What trends do you see emerging in the autonomous vehicle industry ?
T. Nguyen: out with respect to recognition, analysis and perception. The arrival of greater computing power. AI requires a demand for high computing power which can be a hindrance to its democratization.
E. Chevrier: Artificial intelligence is still a black box. We are not able to understand how decisions are made. This poses a particular problem in aeronautics and automobiles, where we have to submit analysis reports of what happened and comply with regulations, including safety. Artificial intelligence is not yet able to find the right solutions alone, but perhaps later, with the right objectives and the right data.
In the context of simulation tools, do you consider AI to be a revolution ? What is your feeling about the impact of the latter in the near future ?
T. Nguyen: I would say that it is not a revolution yet. That said, it brings a lot of potential, opportunities, difficulties and questions. Despite the announcement effects of some, we are still exploring all that this can bring in a practical way.
A word of conclusion ?
E. Chevrier: In the field of AI, we have research projects with IRT SystemX. We are also involved in the government initiative Pack IA and are in contact with the company Quantmetry. With their help, we will launch an internal pilot project to further familiarize our engineers with AI and define a roadmap for introducing AI into our software products. As such, we will identify and prioritize use cases and list the topics on which to introduce an additional dose of AI will be a real plus for our customers.