Background:

Understanding and forecasting the dynamic evolution of complex driving environments is a core challenge in autonomous systems. Traditional object-centric approaches (e.g., bounding boxes) have limitations in capturing fine-grained scene structure and future dynamics. Occupancy representations provide a dense, grid-based description of the environment and enable fine-grained modeling of both static and dynamic elements, forming a solid foundation for world models in autonomous driving. Recent research has demonstrated the effectiveness of occupancy-based world models for future scene prediction. OccWorld introduces a generative 3D occupancy world model that jointly forecasts scene evolution and ego motion via tokenized occupancy representations and transformer-based temporal modeling. UniOcc further provides a unified benchmark and evaluation protocol for both occupancy prediction and forecasting, including support for cooperative driving scenarios and per-voxel motion annotations. Together, these works establish a strong methodological and benchmarking basis for next-generation occupancy world models.

Your Tasks:

• Review and analyze state-of-the-art methods for occupancy world modeling
• Evaluate SOTA models under unified benchmarks
• Design and implement a multi-agent OccWorld, enabling collaborative occupancy forecasting and collaborative path planning
• Evaluate performance and analyze the impact of communication and fusion strategies.

Your Profile:

• Background in machine learning, robotics, or computer vision
• Solid programming skills (Python, PyTorch)
• Interest in autonomous driving and multi-agent systems
• Independent, research-oriented mindset and the ability to explore open-ended problems

Start date: Immediately

Duration: As per the applicable examination regulations.

If you are interested or have any questions regarding this thesis position, feel free to contact.