AI POWER SIMULATION AT DFKI - ASCS Brochure #2025 - Magazine - Page 28
AI POWER SIMULATION AT DFKI
AI-POWERED SIMULATION METHODS
FOR AUTOMATED DRIVING AND
INDUSTRIAL MANUFACTURING
Artificial intelligence is fundamentally transforming
simulation in the automotive industry. The German
Research Center for Artificial Intelligence (DFKI) is
developing modular, data-driven simulation methods
that are used in both the creation of automated driving
functions and the assurance of quality in manufacturing,
with the goal of enhancing safety, efficiency, and quality.
REALISTIC SIMULATION FOR AUTOMATED
DRIVING
DFKI employs a comprehensive simulation approach to
develop safe automated driving functions. Key components
include digitally reconstructing urban scenes, utilizing
generative methods to create scenarios, and realistically
modeling human behavior.
A central element is the automated, digital reconstruction of
real traffic environments. High-resolution, semantically
enriched 3D models of urban infrastructure are generated
from point cloud data, such as that collected by mobile
mapping systems. The underlying technology enables the
efficient processing of large data volumes in terms of
storage, even when the data is very sparse, such as in
fine-grained voxel grids. This precise digital twin forms the
basis for reliable simulations and subsequent AI processes.
Based on this, DFKI is creating generative AI processes for
automatic scenario generation. Large language models
(LLMs) combine text descriptions with structured knowledge
to produce complex traffic situations.
28
A specially built interface translates these into common
simulation formats. Built-in self-validation ensures the
consistency and plausibility of the generated content.
Another focus is on realistically modeling human behavior in
road traffic. Imitation-based learning utilizes expert
trajectories to differentiate between safe and unsafe
navigation strategies. This is enhanced by neural networks
inspired by human visual processing, including graph-based
architectures that model visual attention and integrate
multimodal data – such as RGB images, depth maps,
semantic segmentation, and skeletal data. In this way,
complex scenes are captured comprehensively, and relevant
information is prioritized.
To fill gaps in sensor data, autoencoder-based models are
used to complete partially obscured or incomplete point
clouds. This improves the database for AI training and allows
the simulation of critical scenarios that would be difficult or
dangerous to capture in real life.
These
interconnected
technologies,
ranging
from
scene-based reconstruction and semantic scenario
generation to behavioral agent simulation and automated
data enrichment, facilitate the creation of an entire AI
simulation chain for automated driving. These processes are
utilized in various industrial collaborations as well as in
publicly funded research projects, for example, in the context
of safety-certifiable AI.
