Thesis - Generative & Foundational Terramechanics ML Models for Road Classification

We are Geely Technology Europe. A unified European R&D centre within Geely Auto Group, where world-class engineers, developers and innovators push the boundaries of intelligent mobility. Our mission is to shape the next generation of vehicle architectures, digital technologies and intelligent systems for global markets. We integrate European customer and regulatory requirements early in the development process and support multiple brands within the Geely portfolio, including Zeekr, Lynk & Co and Geely. With nearly two decades of engineering experience in Europe, we continue to build smart, sustainable and user-centric mobility solutions.

Project Description:

Vehicle motion and safety functions perform better when they know what road surface the vehicle is on. The road surface can be classified from signals the vehicle already measures, using supervised time-series classification models trained on labeled recorded drives. This thesis builds on a previous master's thesis successfully carried out in our department and extends that work in three directions. First, some road surfaces are rare, hard to reach or costly to collect data on, so data augmentation and generative modeling on existing recordings could add road surface classes for them and improve the data balance. Second, data from several vehicle models on similar road surfaces exists, which opens the possibility of one pre-trained model whose learned representations generalize across vehicle variants despite the domain shift caused by their mechanical architecture. Third, today's models use a fixed input sequence length (context window), while some road surfaces benefit from a longer one, so a context length chosen at inference time could serve each road surface better.

Scope

The thesis investigates three extensions to road surface classification and evaluates them against the existing approach.

• Data analysis and baseline: characterize the class and domain distributions of the available data across road surfaces and vehicle models, and establish a baseline, evaluation metrics and train/validation/test splits.

• Synthetic data generation techniques: investigate data augmentation and generative modeling methods, such as GANs and other structured and unstructured style transfer techniques, to produce realistic additional samples from existing recordings, to add road surface classes that are hard to collect data for and improve the data balance, and assess both the fidelity of the generated data and its effect on per-class classification performance.

• Towards foundational cross-vehicle generalization: investigate whether a single generalizable road surface classification model can be pre-trained on data from multiple vehicle models, and how well it transfers under domain shift to vehicles with different platforms.

• Variable context length: investigate sequence models that accept variable-length input, so the context length can be chosen at inference time, and study the trade-off between context length, classification accuracy and inference latency for different road surfaces.

• Real-time feasibility: assess model size, computational cost and inference latency with a view to embedded use.

Expected Outcomes

The thesis is expected to show how far synthetic data can extend road surface classification to hard-to-collect road surfaces, how well one model generalizes across vehicles, and what a variable context length gains, within real-time limits.

Deliverables

• A quantified baseline of current road surface classification performance per road surface class and per vehicle model.

• A method for generating synthetic road surface data from existing recordings, with an evaluation of its fidelity and its benefit for adding hard-to-collect road surface classes and improving the data balance.

• An assessment of the feasibility of a single road surface classification model that generalizes across vehicle models, compared with vehicle-specific models.

• An approach for a variable context window whose length is chosen at inference time, with an analysis of the accuracy-latency trade-off.

• An evaluation on recorded data and a real-time feasibility assessment of the proposed methods.

• A master's thesis report, with the aim of a scientific publication.

Your skills and background

• Ongoing Master's degree in Data Science, Machine Learning, Complex Adaptive Systems, Vehicle Engineering, Systems and Control, or a related field
• Strong background in machine learning and deep learning, including time-series or sequence modeling
• Experience with one or more generative modeling, data augmentation, representation or transfer learning, and domain adaptation
• Experience with machine learning frameworks (e.g. PyTorch, TensorFlow)
• Knowledge in signal processing, vehicle dynamics, control systems or applied physics
• Strong programming skills in Python; MATLAB/Simulink and C/C++
• Familiarity with embedded or real-time computing systems
• Strong analytical and problem-solving skills

Why you should join Geely Tech Eu

We are engineers, developers, and innovators from around the world. Joined together by entrepreneurship, our unique blend of global culture, and a belief in a smarter more sustainable future. At Geely Tech Eu we fast-track innovation and transform ideas into pioneering technology solutions, doing your master thesis here is no different. We are convinced that a thesis project is a major contribution to our innovation capabilities and long-term development. You'll have a great opportunity to use your skills and creativity to push the boundaries of what's possible.

What happens when you apply
If this sounds interesting and you match the requirements, please don't hesitate to submit your application with a CV and cover letter. Shortlisted candidates will be contacted for an interview to further discuss the project's details and expectations.

This thesis project is intended for 1-2 students. Applicants may name a preferred partner in their cover letter and must attach a university transcript.

Supervisor: Karthik Prasad, Expert Motion Systems, karthik.prasad@zeekrtech.eu , please contact for more information about the project

Starting date: January 2027

Last application date: 2026-10-31

Apply today. We will perform ongoing selection during the application period. We look forward to hearing from you!

Please note that due to GDPR regulations, we can only accept applications sent through the recruitment system, not via email or other channels.