Motivation
The CARIAD-Bosch Automated Driving Alliance is developing next-generation ADAS/AD systems that require advanced driver behavior models for scalable simulation environments. Traditional approaches using sensor data playback cannot respond to the ego vehicle behavioral changes. Rule-based agents require large engineering effort, lack flexibility, and still have a large sim-to-real gap compared to recorded human behavior. For scaling out our development in autonomous driving, we need robust behavioral models that can generate realistic, interactive agent behaviors at scale.
With the recent development in data-driven behavioral models, reinforcement learning, transformer-based architectures and token-based trajectory prediction models have shown great promise in generating realistic agent behaviors, motivating the exploration of their potential in enhancing simulation fidelity in production systems.

Goals
This project aims to develop data-driven behavioral models that enhance driver simulation realism and scalability for our ADAS/AD development pipeline. The main objective is to create robust simulation agents that realistically react to ego vehicle actions while maintaining kinematic guarantees and handling large-scale scenario generation.
Key tasks:

• Investigation of data-driven behavioral models for sim agents, including next token prediction trajectory models.
• Training and evaluation of the behavioral model on the motion trajectory dataset.
• Hybrid model development combining pure transformer-based models and physical/traffic rules
• Improvement and rule augmentation of existing transformer-based behavioral model.
• Implementation and design of next token-based trajectory prediction models.
  

Methodology

• Literature review of state-of-the-art data-driven behavioral models for autonomous driving simulation, with the selection of a few methods to replicate and extend.
• Data preprocessing to prepare training and evaluation pipelines.
• Model development using the PyTorch deep learning framework to implement and train the behavioral models.
• Utilize and/or extend existing reinforcement learning pipelines for training and evaluation.
• (Optional) Development of additional evaluation metrics to assess the realism of the behavioral models.
  

Data
The project will utilize a comprehensive motion trajectory dataset collected from real-world driving scenarios, including hundreds of thousands of extracted trajectories from more than 1,500 hours of driving data as a baseline. Additional datasets may be incorporated as needed to enhance model training and evaluation.

Requirements and Opportunities
This project offers a unique opportunity to work on innovative problems in autonomous driving for the CARIAD-Bosch Automated Driving Alliance. You will contribute to developing state-of-the-art behavioral models that have a direct impact on next-generation ADAS/AD systems. This project is an excellent fit for students with a strong foundation in at least one of the following domains:

• Academic coursework or project experience in deep learning, reinforcement learning, time-series forecasting, NLP, or computer vision.
• A solid understanding of mathematics, including linear algebra, probability, geometry, statistics, and optimization.
• Strong software engineering skills, particularly with Python and its machine learning-related libraries (e.g., PyTorch, JAX).

Apply to this project here