Efficient and Generalizable World Models for Real-Time Planning in Autonomous Heavy Duty Trucks

Purpose and goal

The project aims to develop efficient and generalizable world models for real-time autonomous driving in heavy-duty trucks. The research focuses on compute-efficient multimodal representations that enable robust trajectory planning under strict real-time constraints while generalizing across varying vehicle configurations, loads, sensing conditions, and driving environments.

Expected effects and result

The project is expected to advance scalable autonomous driving technology for heavy-duty vehicles through improved real-time planning performance and computational efficiency. Results include new methods for compact multimodal world models, improved robustness under changing vehicle dynamics, and evaluation frameworks tailored to articulated trucks. The project contributes to safer, more energy-efficient, and competitive transport systems.

Planned approach and implementation

The project combines industrial research at TRATON with academic research at KTH to develop compute-efficient world-model architectures for autonomous heavy-duty trucks. The work includes multimodal representation learning, robust trajectory planning under varying vehicle dynamics, and development of evaluation frameworks tailored to heavy-duty trucks. Methods will be validated under realistic deployment conditions and real-time computational constraints.