ATTi talks to Kia Cammaerts of Ansible Motion to find out how state-of-the-art simulators are being used to advance vehicle, tire and powertrain testing programs.
How are simulators useful for powertrain calibration?
You could run offline simulations with software models of EV powertrains or ICE powertrains, you can run them faster than real time and run hundreds of thousands or millions of them and derive significant information about the whole gamut of behaviour. Then you might add more detailed models, and at some point, you’d want to add hardware-in-the-loop, running real ECUs and real code, but [the disadvantage is] you can’t speed them up, or slow them down. To get exactly the same responses [as they would in reality] they have to run in real time.
Equally you might want to put a driver-in-the-loop, there is no limit. You can also build on this and add mechanical hardware-in-the-loop, like significant parts of the physical vehicle. At the extreme end, [you can use] a real human, real engine, real gearbox, real braking system, real cooling system, and a virtual vehicle. At the offline end, everything is virtual and can be run faster than real time.
You can certainly do the majority of [powertrain] simulation without mechanical hardware, electronic hardware, or human wetware-in-the-loop, because you can explore very detailed mathematical models of the behavior of your system in a great degree of complexity. But at some point, you do need to focus on in-the-loop experiments.
Mechanical hardware-in-the-loop is rare, because of the complexity and the expense, but it is the gold standard for simulation. The big disadvantage is that you actually have to have the hardware there. If you have a good software model of the hardware you are proposing, you can put a human in at any stage, and you can actually put them in at a much earlier stage than you can with HiL or MHIL applications.
In what ways are simulators or DiL being used for ADAS or AV testing?
The offline simulation of ADAS is not mature, but it has been going on for a long time. The big trend is improving the quality of the simulation of sensors and the sensor’s perception of the external world.
Over time, we’ve seen improvements of the complexity of sensor models, especially for HiL testing where you’re actually using real-world hardware. You have to be able to model the sensor aspects that are under examination in real time. [For example], really brilliant microwave sensor models take into account loaded and instant radiation scattered from the atmosphere and secondary or tertiary reflections, and cross echoes from other vehicles. It’s not currently possible to run those effectively at the highest level in real time so you can’t use them in HiL simulators.
If you put humans into ADAS-based scenarios, you lose the ability to run them faster than real time and therefore lose experimental capacity [because you don’t have] a hundred humans that behave identically, all running at the same time. What you do gain though is human aspects that cannot be simulated.
Classic tests [scenarios] are handover, and handover from human to machine. It could be the pressing of the automatic drive button, or it could be a less obvious situation like falling unconscious at the wheel. That transition is an important one, but even more important perhaps is the transition from machine driving to human driving, because if I am semi-conscious, staring ahead but not really taking information in, the machine may decide that the conditions ahead are tricky and it would much rather the driver takes over.
That is a point of maximal danger [and] is a focus of a lot of experimentation at the moment. It will remain a problematic point, the handover from the machine to human, and it will be problematic as long as there is a human to hand over to.
As ADAS improves, as it gets more complex and we move toward Level 3 and Level 4, we’re starting to realize aspects of the dream of autonomous driving without fully autonomous Level 5 vehicles, and without Level 4 being available for the whole journey. We are still starting to feel the benefits of some of those offers that were made a while ago but have not really materialized yet.
We are moving beyond just putting a driver in the loop, adjusting ADAS or AV technology so that it is able to drive without crashing, but being able to drive in a way that is perceived as good by a human.
We are going to go beyond that, to it adapting its style to how the human is reacting to it, which you definitely need a DiL to detect, and then we will go one stage beyond that, where we will have brand identity coming out, where a harder edge more aggressive sporting car, from a prestige German background, will have a style of driving that is different to a more city- or commodity-oriented car from an American background.
How are simulators useful for developing and testing tires, and have you seen a growing interest from tire manufacturers?
The fundamental problem with simulating tires is the massive complexity of tire behaviour in the real world. When driving a car, all forces acting on your body come from the tire’s interaction with the ground.
In a good braking or steering system, you get very direct feedback of the tire performance, but all parts of the driving experience dynamically come from tire interactions with the road, so it’s not difficult to do a simulation of a tire that captures the majority of the performance attributes, for example, like peak grip or primary ride or basic slip angle characteristics and basic understeer/oversteer. A great deal of the human’s perception of the driving experience comes from the bits that are very difficult to model.
This comes back to the whole human as a supercomputer and humans taking in all information and synthesizing an opinion out of it. It’s very hard to model so humans will form a nuanced opinion of how tire contact patch forces are being fed through the chassis, into the steering and into the seat, but they will form an opinion in seconds of getting into a car. As you drive out onto the proving ground, or a track, you will already have an opinion about that car. This is very hard to simulate, and also very hard to objectively quantify even if you are simulating it.
One challenge is bringing greater nuance into the detailed simulation of tire contact behaviour, another problem is quantifying the quality of the human experience of that. The third problem is if you want to put human-in-the-loop or driver-in-the-loop, you have to bring your models under the constraint of being able to solve in real time.
The very top tire models cannot solve in real time as they take longer than a second to simulate a second of behavior. There is a problem of simplifying, but we have seen recently that tire models that operate in real time are capable of reproducing highly nuanced behavior and performance attributes of those tires, to the extent that OEMs are able to develop vehicles around tire models, and also tire manufacturers can start to develop tires which can be modeled to suit OEM preferences.
Another thing we have seen is the much more aggressive use of tire models as a method to exchange performance data between Tier 1s and OEMs. This includes Tier 1s buying high-end simulators, and OEMs pressing Tier 1s to improve the model of their tires for human-in-the-loop and driver-in-the-loop purposes. This is distinct to the top end of tire modeling, which is offline, slowed or in real time, from which you can get a lot of data. Vehicles are now being developed with a much higher confidence about tire performance, and tires are being selected on the quality of their performance in DiL simulators.
Can you use simulation for crash testing?
DiL simulation is a great test environment for [crash testing] because you can put a human into a very dangerous situation and see if – by using their car control or the vehicle’s active safety – an accident is avoidable. However, we still need a human to simulate what a human would do in these difficult situations.
Avoiding the accident is an excellent use of DiL – you can have situations where accidents are the outcome, but you are not simulating the point of impact, [instead] you are simulating everything up to that point. Was it a bad outcome for the occupant, or was it a bad outcome for the other road user for example?