Internship - Enhancing a Diffusion-Based Traffic Scene Generation Model for Closed-Loop Simulation - REF5865M

High-fidelity traffic scenario generation is increasingly important for developing and validating autonomous driving systems. Diffusion models have recently demonstrated strong potential for synthesizing diverse and realistic multi-agent driving scenes.

In this internship, you will contribute to advancing an existing diffusion-based traffic scene generation model used in closed-loop simulation within the Aumovio AI Lab in Berlin. The focus of your work will be on improving the model's generation quality through enhanced training strategies, better data representations, and the development of a critic module for automated quality assessment.

You will have direct access to our simulation stack, enabling interactive testing and evaluation of your improved models across a wide range of simulated driving scenarios.

Duration: 6 Months

Tasks:

• Conduct a familiarization phase and baseline analysis of the current diffusion model.
• Enhance the existing diffusion model using improved conditioning, data augmentation, fine-tuning, or reinforcement learning methods.
• Develop a critic model capable of assessing scene realism, safety, and consistency.
• Integrate the improved generator and critic into the simulation framework and conduct a comprehensive evaluation in closed-loop settings.