Low-speed AD applications offer many potentials for a wide range of stakeholders. E.g. they represent a potential path for manufacturers and regulators to introduce highly automated driving functions (SAE Level 3 and 4) into the existing traffic, as they pose limited risk due to the low kinetic energy. In addition, the technical complexity for perception, prediction, and planning is reduced. The most popular and technically advanced use case is Automated Valet Parking (AVP), which is already found in industrial series and regulatory approved since 2022. An expansion of these possibilities is the application of low-speed AD functions in public road traffic. It is to assume that dedicated low-speed functions can take over the driving task in specific low-speed situations and thereby supplement capabilities of the full-range function or replace them with a more energy-efficient operation. In current research and state-of-the-art technology, these use cases are mostly realized by independent driving functions, tailored to each specific application. To efficiently combine them in one AV architecture, a holistic approach of a generalized low-speed function is of interest. To close this gap, one goal in the AUTOtech.agil project is to develop a generalized AD function for low speeds (“Generalized Low-Speed Function”, GLSF) based on one superordinate functional architecture. This GLSF comprises its own sensor module and software stack and shall cover as many different use cases in the low-speed range as possible. Thereby, the project builds on the experience gained from the function “Safe Halt” of the project UNICARagil, which is also a low-speed function with proprietary hard- and software.

This topic is researched by Moritz Berghöfer.

Within the UNICARagil project, four highly automated vehicles were developed, which have sensor modules for environment perception. For the automated driving function, an environment representation model is created from the recorded sensor data using perception algorithms. For perception, the recorded data is processed on GPUs by artificial neural networks (ANN). This form of data processing is very energy intensive due to the amount of data and the type of processing.

In AUTOtech.agil, therefore, methods are now being researched to enable more energy-efficient perception through data processing on neuromorphic hardware. So called spiking neural networks (SNN) are used, which can be operated particularly efficiently on neuromorphic hardware.

This topic is researched by Lorenz Bayerlein.

Automated driving offers the opportunity to improve road safety. Nevertheless, proving a proof that possible incorrect behavior of automated vehicles does not pose an additional risk to road users is one of the key challenges in the introduction of such vehicles, as human drivers no longer act as a fallback level. Due to the complex environment in road traffic and the increasingly complex systems in the vehicle, established safeguarding methods result in efforts in the order of 10^9 kilometers. The approach of UNICARagil was to secure individual modules in isolation instead of the entire system, which both reduces the initial effort and prevents systems from having to be retested again and again after changes to modules. This approach will now be followed up by describing the capabilities of individual software and hardware modules in a “quality vector” and aggregating them into an overall function. In AUTOtech.agil, methods for modular validation are being developed, test cases defined and transferred to a homologation process.

This topic is researched by Alexander Blödel

In automated driving, artificial intelligence (AI) plays an important role, particularly in the area of perception. The main challenge of AI algorithms consists of the generalization quality. If the learning leads to a poor level of generalization, the AI algorithm faces problems with data that is not represented by the training data set. This may become difficult for the occurrence of so called “corner cases”, which describes infrequent and critical scenarios that can only be covered limitedly by the dataset. In the project “KI-Absicherung”, a safety argumentation for the AI-function of pedestrian detection was developed. In Autotech.agil, the assurance strategy is being expanded to further low speed functions. Furthermore, a uniform and system wide specification of the AI functions, including an aggregation between modules with AI- and non-AI-functions, is being made in order to develop a modular safety approval.

This topic is researched by Anton Kuznietsov