Neuromorphic hardware for autonomous driving

Mr. Rothe, you coordinate the KI-FLEX project as a project manager at Fraunhofer IIS. What is the aim of the project?

There’s an urgent need for technology components that make autonomous driving safer and more reliable. In the KI-FLEX project, we’re rising to this challenge and developing a hardware platform that helps to measure vehicle position and determine vehicle environment precisely using artificial intelligence. The key thing about this platform is its flexibility. This is essential because, whereas the product cycles for vehicles are long, artificial intelligence is advancing rapidly. We’re developing the platform so that future developments in artificial intelligence and neural networks can be taken into account without the need for complete redevelopment. The platform is therefore software-programmable and reconfigurable.

 

What lies at the heart of this hardware platform?

The core component is a hardware chip known as an application-specific integrated circuit (ASIC), which is combined with a field-programmable gate array (FPGA) to speed up the applied AI algorithms. The key strength of the ASIC is its excellent power efficiency, although the development costs are very high. We’re therefore bringing together the enormous flexibility of an FGPA with the exceptional power efficiency of the ASIC – as well as other flexible elements – in order to combine the advantages of both systems. Another special feature of the chips is their highly energy-efficient operation – which is essential if artificial intelligence is to be applied directly in the car. Hardware of this kind, which is specifically designed to accelerate neural networks, is also known as neuromorphic hardware – because it’s inspired by neural systems in biology.

 

 

Flexibility is essential. What exactly does that mean?

We use various sensors to record the real traffic conditions, both static and dynamic. A specific neural network (NN) is then used to analyze these data – and, for example, to initiate a braking process if a child runs out into the road. The aim is to be able not only to work with the existing neural network, but also to use the platform to implement future improvements in NN architectures, thereby keeping the platform constantly up-to-date with the latest developments in artificial intelligence.

 

What fascinates you about this project?

The fascinating thing is that we’re operating in not one but two cutting-edge fields that are increasingly gaining traction with the general public: autonomous driving and artificial intelligence.

 

What is your role in the project?

KI-FLEX is a joint project involving Fraunhofer IIS and seven other partners from research and industry. As Fraunhofer IIS is responsible for coordinating the project, I bring together all the different activities as well as leading the subprojects undertaken by Fraunhofer IIS. These primarily revolve around developing a flexible, digital accelerator core and integrating both this and other accelerator cores into the chip (ASIC) that is being developed as part of the project. These accelerators are instrumental in allowing artificial intelligence to be applied directly in the car.

 

How far is it from the current state of research to getting “out on the road” – in other words, to incorporating the platform into self-driving vehicles?

The project runs until the end of August 2022, so we’re still in the development phase at the moment. In the last step of the project, the plan is to integrate the platform into a vehicle to create what is essentially a proof-of-concept. By doing so, we want to demonstrate that our technology works as planned.

Links:
https://www.iis.fraunhofer.de/de/ff/kom/ki/neuromorphic/ki-flex.html
https://www.iis.fraunhofer.de/de/pr/2019/20191203_ki-flex.html