Automated driving functions are still significantly limited in their scope of use. The reasons lie in the system architecture currently employed and the discriminative methods of machine learning used. Based on generative methods, NXT GEN AI METHODS introduces bidirectional information flow as a new paradigm into the operational chain, enabling massive improvements in the development of autonomous driving functions.
Specifically, this leads to better scalability through an inexhaustible reservoir of data for offline testing, validation, training, and online error detection; better transferability through the ability to deconstruct and recombine semantic information and expand the Operational Design Domain (ODD) by generating targeted scenarios and sensor data; and better traceability through online verification, validating each processing step in the operational chain during operation, as well as understanding the latent space. A key outcome for industry will be the development of foundation models for driving data.
The Munich University of Applied Sciences (HM) is researching generative models for generating lidar and camera data in complex urban scenes and expanding these with consistent fusion in the latent space. Additionally, HM is exploring sequence models for scene prediction and vehicle trajectory planning, focusing on predicting the trajectories of vulnerable road users and modeling uncertainty. Furthermore, HM is working on transferring foundation models for sensor data generation and scene understanding in the automotive sector.
The project is being conducted at the Intelligent Vehicles Lab (Department of Electrical Engineering and Information Technology, Institute for Applications of Machine Learning and Intelligent Systems – IAMLIS) under the guidance of Prof. Dr. Fabian Flohr.
More information on: https://nxtaim.de/