Impulse Project 13: Hybrid Models and AI Methods for Safe Mobility (HyMne 2) - Data Generation and Data Quality

The SAFIR impulse project “HyMne2” investigated the combination of domain knowledge in vehicle dynamics with machine learning methods. The aim was:

1. to improve state estimation,
2. to generate and control realistic driving behaviour, and
3. to identify relevant traffic scenarios for the validation of automated driving functions.

The project, led by Prof. Dr.-Ing. Michael Botsch, comprised three sub-projects:

AI-based optimisation of correction data for inertial measurement systems

Highly accurate vehicle state estimation is essential for the development and validation of automated driving functions. This is usually based on a combination of inertial navigation systems (INS) and global navigation satellite systems (GNSS). However, in environments such as tunnels or multi-storey car parks, GNSS availability is limited, which reduces estimation accuracy. Additional sources of correction or reference information are therefore required. 

In Sub-project I, on-board sensor data was used to generate correction data for the INS, in particular for estimating the yaw angle. To this end, a novel uncertainty-aware hybrid learning model was developed, which combines model-based approaches with a Transformer architecture whilst taking uncertainties into account.

The training data was sourced from real-world tests at the CARISSMA open-air test site and has been made publicly available to the research community. Integration into an Extended Kalman Filter and real-time implementation on an NVIDIA ORIN platform were successfully validated.

The industry partner in this sub-project was GeneSys Elektronik GmbH.

Guided Driver Behaviour Generation

In Sub-project II, a hybrid model architecture was developed that enables the targeted manipulation of data-driven driver behaviour models. This is particularly relevant for simulations, in order to specifically influence the realistic driving and movement patterns of road users and thereby achieve realistic movement patterns.

The AI model proposed for this purpose in the sub-project utilises historical driving data and ‘guidance’ information to derive control variables such as longitudinal acceleration and yaw rate. A downstream physical motion model transforms these signals into consistent and realistic vehicle trajectories, thereby combining learning-based behaviour modelling with physical realism.

Two AI architectures were validated in the sub-project: an LSTM-based architecture and a Transformer-based architecture. As a physics-informed generative model, the latter enables the generation of multiple plausible trajectories under the same boundary conditions.

A key challenge regarding trainability was resolved through the use of Gumbel-Softmax sampling, which enables efficient training despite the presence of non-differentiable components. This methodological approach can be applied to other comparable AI model architectures.

The industry project partner in this sub-project was ZF Mobility Solutions GmbH.

Selection of relevant scenarios for the validation of driving functions 

The testing and approval process for automated driving functions requires suitable and relevant test scenarios in scenario-based testing. In Sub-project III, methods were therefore developed to identify critical as well as rare or atypical traffic situations.

Firstly, existing criticality criteria were expanded to enable the rule-based detection of critical scenarios. Subsequently, scenarios from our own dataset and from the INTERACTION dataset were converted into a latent representation using metric learning and clustered on this basis. Particular focus was placed on outliers, i.e. scenarios outside the formed clusters.

The methods were applied to real-world traffic data from an inner-city junction in Ingolstadt, as well as to specifically generated rare scenarios recorded at the CARISSMA open-air test site. This demonstrated that the developed methods are suitable for identifying relevant traffic scenarios.

The industry project partner in this sub-project was Audi AG.

Result: HyMne2 provides innovative approaches to improving state estimation, realistically simulating driving behaviour and identifying relevant test scenarios – key building blocks for the safe development of automated driving functions.