Discover more about the topics and technologies to be discussed at this year's (fee-to-attend) conferences, via a series of exclusive interviews with a selection of our expert speakers
Gianluca Vitale, global business segment manager for AVL List, previews his presentation on the company’s alternative to brute force simulation testing and explains why simulation should support real-world testing, rather than vice versa.
Catch Gianluca’s presentation HPC, artificial intelligence, virtualization – a creative combination for ADAS development at the Autonomous Vehicle Software & AI Symposium. Purchase your delegate pass here.
Tell us more about your presentation.
We have been working with AI in powertrain development for a good 20 years now, and we are now starting to apply it to new areas. The need for AI in powertrain development is because of the extreme complexity of the control units – I’d say they’re one of the most complex systems on the market, so we are very well placed to apply our knowledge to autonomous driving.
Combining our expertise in simulation with our experience in virtualization, we wanted to see if we could offer a valid proposition for ADAS testing, making the process more cost-effective. With access to high-performance computing, advanced AI, and componentized software, we decided to set up a research program to find a better way to do things.
The basic idea of our approach is one of efficiency. A traditional brute force approach, where you scout for each and every possible combination of parameters for a scenario, is a waste of time, because you know that certain combinations are not critical at all in real life. For example, when a car approaches another at slow speed and it is still a long way away, why bother simulating and testing it? You should go and scout for the scenarios that are at the limit of safety. This is what we’re doing, using reinforcement learning.
How can you trust that the AI doesn’t leave out critical situations?
Every scenario can be described by a set of parameters. This results in a certain ‘design space’, an n-dimensional environment where all the cases are included. You can be quite sure you will hit the nail on the head because you’re not leaving a certain criterion area. Instead, you can simply eliminate the situations that are guaranteed not to be relevant. Of course, this does require the person who is designing the optimization to input the correct values.
Can you quantify the benefits of such an approach?
We have calculated that normally – with all the different possibilities of vehicle variance, calibration and environmental conditions – it would take 660 years in a real-time environment to go through all the combinations necessary to validate an ADAS software package. Of course, simulation is often slower than real-time.
Thanks to reinforced learning and the use of multiple computers, the simulation and AI process can be parallelized and scaled on a server (up to a factor of 1,000), distributing it over a large number of cores. This way, we can bring those 660 years down to just about a week.
What is the scope of this technology?
I want to make clear that we don’t envision a validation process consisting purely of simulation. Our idea is to try to make the in-vehicle testing phase as efficient as possible and reduce wasted time. The intention is not to identify the scenario, run the simulation and call it a day. Instead, we want to identify the real scenarios that the testers in the car should focus on.
It’s always been our aim to use simulation as much as possible, but in the end, you can never do away with the real-world validation. There are things that simply can’t be done just by using simulation. At some point, after the simulation hype dies down, we’ll arrive at a good compromise, where we’ll do as much as possible in simulation, but always leave something for the real-world testing.
The result will be more intensive usage of real data, of which there will be more and more, as we’re starting to get access to more in-use data from cars being run by the public. This kind of data is precious to us as it will help to reduce the amount of vehicle prototypes that get built.
There are a lot of people playing with autonomous technology, but in the end, there are only a handful of technologies that really have an innovation impact. You always need to ask the question of the benefit to the customer. That is why, rather than aiming for a completely virtual development environment, we focus on the here and now, where a lot of the testing is done in the car. Except, what we do is to make that testing more effective, saving time.
What is the roadmap for this technology?
We think that the complete solution will be ready in around two years. One of the things we are still working on is making the technology work with the cloud to enable automatic running and distributed resources. We have a prototype for this, but currently, the solution runs on workstations. The next thing we’ll do is parallelize all the software processes to be able to use them on a high-performance computer so that we can parallelize tasks.