Open Project Day

Using Generative Models for Automated Driving, Joint AI Innovations by Academia and Industry

E-Booklet Open Project Day

Using Generative Models for Automated Driving, Joint AI Innovations by Academia and Industry

Open Project Day with nxtAIM highlight-, deep dive-, and poster presentations at the University of Freiburg

The nxtAIM Open Project Day, held at the University of Freiburg on February 25, 2026, brought together project partners, experts from academia and industry, and members of the funding body to showcase the joint progress of the consortium. With around 120 participants, the event fostered lively exchanges with nxtAIM experts and provided an in-depth look at current research directions and project achievements.

Central Topics and Project Highlights

A key focus of the project lies in advancing generative AI methods. Instead of training functions directly from data, the nxtAIM consortium models the data itself, creating the framework for versatile foundation models, thereby paving the way for versatile foundation models in automated driving.

In the area of training infrastructure, data, and collaborative development, nxtAIM achieved significant milestones in 2025 that were central to the project’s technological progress. These included the successful acquisition of computing resources at Europe’s first exascale supercomputer JUPITER, enabling largescale model development.

The required data are provided through a concept established in nxtAIM, enabling the GDPR-compliant use of proprietary data from multiple partners. Together, these accomplishments laid the foundation for coordinated data handling, shared development pipelines, and the training of generative models tailored specifically to automated driving. These method-driven contributions align closely with broader European efforts to promote federated training and open, industry-wide collaboration, approaches that are already deeply embedded in nxtAIM’s work.

Highlight Presentations and Guest Speakers

The day was enriched by a keynote from Prof. Andreas Geiger (University of Tübingen), who offered insights into current research trends in AI-driven perception and generative modeling. His talk provided a strong insight to „The Role of Synthetic Data for Autonomous Intelligence”. Guests were welcomed by Dr. Ernst Stöckl-Pukall (BMWE), who emphasized the importance of collaborative AI research initiatives such as nxtAIM for Germany’s and Europe’s technological competitiveness, virtually welcomed by Ms. Hildegard Müller, President of the VDA and by the nxtAIM coordinators Dr. Jörg Reichardt (Aumovio) and Dr. Julian Schmidt (Mercedes-Benz).

 

Engaging Talks and Poster Sessions

The program included highlight presentations on the consortium’s latest achievements, deep‑dive sessions with lively discussions, and a poster session that encouraged networking among researchers and industry.

Beyond engineering achievements, nxtAIM demonstrated a strong international research presence. Project partners presented their results at major conferences across Europe, the United States, Korea, and Australia, like the British Machine Vision Conference (BMVC), International Conference on Computer Vision (ICCV), Conference on Computer Vision and Pattern Recognition (CVPR), Winter Conference on Applications of Computer Vision (WACV), International Automated Vehicle Validation Conference (IAVVC), and The Conference on Robot Learning (CoRL).

What’s next for nxtAIM?

Looking ahead, the project is entering its final phase. The next major milestone will be the final event in Frankfurt in November 2026, where the consortium will present real-world demonstrations and the completed nxtAIM results.

Virtual Welcome by Hildegard Müller, President of the German Association of the Automotive Industry

Posters – Overview of nxtAIM research results

List of Posters and Presenters

SP 1 Single Frame Multi Sensor

1.1 | Guided Diffusion for Scalable and Semantically Controlled Augmentation of Automotive Camera Data Carolin Waltraud Wunderlich

1.2 | Image-Conditioned Radar Generation for Multimodal generation Jonas Neuhoefer

1.3 | Adapt or Augment: How to Utilize Generative Models? Joshua Niemeijer

1.4 | Synthetic FMCW Radar Range-Azimuth Map Generation Christopher Grimm, Claas Tebrügge

1.5 | Extending Automotive Camera Field of View with Generative Outpainting Omar Alnaseri

1.6 | TREAD: Token Routing for Efficient Architecture-agnostic Diffusion Training Ulrich Prestel

1.7 | Style Swap with Multi LoRA Conditioning Malte Ganzer

1.9 | Lidar NeRF, Dominik Scheuble Dominik Scheuble

1.10 | Clustering-Based Dataset Curation Frithjof Marquardt

1.11 | Evaluation of Conditioning Modes for AI Image Generation Sven Burdorf, Meghna Prabhu

1.12 | Improving Synthetic Data Quality through Data Pruning Leroy Odunlami, Sven Burdorf

1.13 | Latent Diffusion Model for sim2real Lidar Data Generation Carolin Waltraud Wunderlich, Vaithiyanathan Alagar

1.14 | Solar Altitude Guided Scene Illumination Samed Dogan

 

SP 2 Multi Frame Single Sensor

2.1 | Criticality-guided Diffusion for Counterfactual Traffic Scenario and Video Generation Adam Molin

2.2 | Self-Supervised Learning for RADAR Claas Tebrügge

2.3 | Scaling Automotive Video generation via Sparse Mixture of Experts Christian Ojeda-Bernal

2.4 | Orbis Extensions: Conditioning, Compositional Generalization, Feature Probing Sudhanshu Mittal

2.6 | Geometry-Aware Latent Compression for 3D Gaussian Driving Scenes Marius Kästingschäfer

 

SP 3 Abstract Scenarios & Planning

3.1 | WorldVLM: Combining World Model Forecasting and VLM Reasoning Katharina Winter

3.2 | Integrating Machine Learning and Control Theory for Safer Motion Planning Bojan Derajic

3.3 | Generating Trajectories with VideoGAN Annajoyce Mariani

3.4 | Divide and Merge: Motion and Semantic Learning in End-to-End Autonomous Driving Ö. Sahin Tas

3.5 | LiFlow: Flow Matching for 3D LiDAR Scene Completion Andrea Mateazzi

3.6 | End2End Autonomouos Driving: Motion Planning Under Distribution Shifts Dikshant Gupta, Michele De Vita

3.7 | Road Graph Generation Laurenz Thiel, Sven Peyinghaus

3.8 | Complete Gen-AI framework from regulations to test scenarios in automated driving validation

3.9 | EP-Diffuser: An Efficient Diffusion Model for Traffic Scene Generation via Polynomial Representations Yue Yao

3.10 | KnUT-Knowing Unkown Things Natalie Grabowsky

3.11 | Self-Supervised Pretraining for Aerial Road Extraction Christian Hubschneider

3.12 | Physics-Based Vehicle Simulation for Open- and Closed-Loop Testing of Generative Trajectory Planning Hendrik Fenske

3.13 | Towards System-Level Evaluation of Driving Functions Christian Schlauch

3.14 | Efficient Generative Rollouts for Autonomous Systems: Map-Constrained Scenario Generation with Hardware-Aware Pruning Shashank Pathak

 

SP 4 Foundation Models

4.1 | PrITTI: Primitive-based Generation of Controllable and Editable 3D Semantic Urban Scenes Christina Ourania Tze

4.2 | Measuring Gradient Noise Marcel Aach

4.3 | Customized LLM for Traffic Scenario and Electric Range Simulation Christian Schyr

4.4 | From Abstract Scenarios to Photorealistic, Synthetic Data Thies de Graaff

 

SP 5 Automotive Scalability

5.1 | Optimizing Generative AI for Integration into Automotive Systems Christian Ojeda-Bernal, Nitin Kannan

5.2 | Low Rank Compression of Neural Networks Arunachalam Thirunavukkarasu

5.3 | Driving Data in the Cloud: OpenDrive Checks & Carla Simulation Workflow Nils Worzyk, Leonard Schroven

Impressions

Scroll to Top