Jake Rankin from Cohort 4 gives us a description of the conference he recently attended which was hosted by ADAS dSPACE and Warwick Manufacturing Group (WMG) at Warwick University entitled 'The State of the Art for Developing and Testing Advanced Drive Assist Systems'.
After an introduction, Dr Phillip Clarke discussed autonomous vehicle testing methods, in particular Hardware in the Loop (HiL) testing. HiL is a key means of being able to increase both the volume and range of testing, working closely with ISO 26262, dSPACE use a combination of Automotive Simulation Models (ASM) and Virtual Validation (Left-Shifting) to help further train simulations. The purpose of the conference was to discuss how this method was implemented and why it was useful.
After discussing the various levels of autonomy, and recognising that several sensors would be needed, three challenges were highlighted:
- Fusing data at different frequencies
- A need for an increasing communication bandwidth
- Validating simulated models
Typically, two methods of testing are used. Open-loop testing is used to check the performance of the algorithm and closed-loop testing is similar but with the addition of a simulation model that gives feedback. HiL is a type of closed loop testing which both records and analyses simulations and feeds real sensors data to see how the simulated models work. Some of the tools included a radar test chamber which contained a radar in a circular chamber with movable rings that emit acoustics to mimic distance. Testing can be done even before a CAN setup is used to just test the algorithms, which is called left-shifting. dSPACE use a system called VEOS to test the model behaviour.
Rapid control prototyping of machine algorithms was the next discussion and firstly focused on the basics of sensor/actuation interaction. In particular, it discussed the means of applying neural networks and deep-learning into an actual system by first developing and training the learner with Tensorflow, with the help of SLAM algorithms. Next, Nvidia hardware such as the Jetson X2 is used to optimise the learning algorithm and the algorithm is finally places into dSPACE’s MicroAutoBox II. The system was shown to be very flexible in terms of coding languages that it can accept as well as other off-shelf programs such as MATLab.
WMG’s Graham Lee then gave an overview of the facilities at Warwick University, part of the CATAPULT group and Innovate UK. Currently, Warwick are finalising their NAIC building which will also give MSc courses on autonomous vehicles; the first in the country. The main project that Graham was working on was a system called SAVVY. This is a project which hopes to develop scalable AI testing, using some of the aforementioned equipment from dSPACE.
Finally, Torten Kluge gave a talk on scenario-based testing and sensor simulation. One of the challenges was acquiring data and dSPACE were able to collect a lot of their data from GIDAS, which are accident studies in Germany. These scenarios would then be uploaded into the simulation, with additional input from services such as SUMO (a traffic-flow simulator). All of this was demonstrated with dSPACE’s simulation which included all of the sensor data, car behaviours and states. Because of the simulation design, even with critical components such as the ECU and sensors, the simulation could still be tested.
Overall, it was a very interesting insight into autonomous vehicles and certainly resulted in useful subjects to research for vehicle autonomy. https://www.dspace.com/en/inc/home/applicationfields/our_solutions_for/driver_assistance_systems.cfm for more information.