Advancing generative AI for autonomous driving with a new simulation framework

Physics-Based Vehicle Simulation for Open- and Closed-Loop Testing of Generative Trajectory Planning

As part of the large‑scale innovation initiative NXT GEN AI METHODS (nxtAIM), IPG Automotive continues to strengthen its role in supporting next‑generation AI technologies for autonomous driving. More than 20 organizations from across the automotive industry – OEMs, Tier‑1 suppliers, technology companies, and research institutes – have joined forces in the three‑year project launched in early 2024 to leverage generative AI for perception, prediction, and planning in automated driving systems.

Within this consortium, IPG Automotive focuses on delivering a simulation framework for testing/validating generative trajectory planning, laying an essential foundation for scalable and reliable AI model development. The recently completed specification phase marks a key milestone in this effort.

A simulation framework designed for the future of generative mobility

The specified framework, built on CarMaker from IPG Automotive, enables project partners to develop and evaluate generative scenario models as well as trajectory prediction and planning algorithms in a unified environment. At the heart of the design are two complementary use cases, capturing the collective needs of all consortium partners:

Open-loop training: Synthetic data at scale

To train AI models effectively, vast quantities of high‑quality data are essential. The simulation framework addresses this by generating large synthetic datasets with Waymo‑style abstract environment representations, including positions of traffic participants, map information, and traffic signal logic. The data is independent of sensor configurations and easily integrates into partners’ Python‑based toolchains – an important requirement for seamless collaboration.

During simulation, all relevant variables are recorded and exported as TFRecord files, making them compatible with widely used machine‑learning workflows. A wide variety of scenario types, including rare and safety‑critical edge cases, ensure meaningful training opportunities for trajectory prediction models.

Closed-loop simulation: Validating generative trajectory planning

The second use case focuses on validating generative models for trajectory planning. Unlike open‑loop workflows, the environment in this setup responds dynamically to the ego vehicle, enabling continuous reassessment of the generated trajectory in each simulation step. Ensuring that identical inputs always produce identical outcomes – strict reproducibility – is a core requirement for the development of robust algorithms.

A gap analysis confirmed that CarMaker already exposes most of the data necessary to support Waymo‑compatible scenario structures, simplifying integration across the consortium.

Modern, scalable, and ready for industrial deployment

To ensure long‑term adaptability, the framework is implemented with full containerization, providing platform independence and easy scalability – both crucial for generating extensive datasets and integrating advanced generative models. This aligns closely with the broader nxtAIM vision of enabling scalable, transferable, and traceable AI systems through generative methods.

Overall, the specified framework creates a robust, standardized, and future-proof basis for training, validation, and scaling of generative traffic scenarios and planning algorithms in the project. By complementing nxtAIM’s cross‑partner efforts in generative modeling, IPG Automotive helps accelerate the creation of future‑ready simulation ecosystems and more capable autonomous vehicle functions.