[Guided Research, Master Thesis] - Open Vocabulary Gaussian Splatting Navigation

27.04.2026, Abschlussarbeiten, Bachelor- und Masterarbeiten

Gaussian Splatting provides a continuous, high-fidelity representation of 3D environments, but lacks an explicit mechanism for semantic reasoning and navigation. While recent extensions enrich Gaussians with feature embeddings, they typically rely on closed-set supervision or remain limited to per-pixel semantics, making them insufficient for open-world interaction.
This thesis proposes an open-vocabulary object-centric representation for semantic navigation in Gaussian Splatting pipelines. The core idea is to move from unstructured Gaussian fields to clustered object-level abstractions derived directly from learned feature embeddings (e.g., CLIP/DINO-aligned features), without predefined class labels. This enables the system to identify, group, and reason about objects in a flexible, task-driven manner.

Current Gaussian Splatting pipelines, such as GSFusion, prioritize geometric consistency and rendering quality, but do not support semantic grouping or navigation-oriented queries. As a result, downstream tasks like object search, goal-directed navigation, or interaction require additional external modules or labeled datasets, limiting scalability in dynamic or unfamiliar environments.

To address this, the proposed framework extends GSFusion with three tightly integrated components:


-Feature-Aligned Gaussian Representation
-Open-Vocabulary Object Clustering
-Semantic Navigation Interface

Required skills:
-Python/C++
-Robotics
-Computer Vision


-Wei, J., & Leutenegger, S. (2024). Gsfusion: Online rgb-d mapping where gaussian splatting meets tsdf fusion. IEEE Robotics and Automation Letters, 9(12), 11865-11872.

-Chen, T., Shorinwa, O., Bruno, J., Swann, A., Yu, J., Zeng, W., ... & Schwager, M. (2025). Splat-nav: Safe real-time robot navigation in gaussian splatting maps. IEEE Transactions on Robotics.

-Chen, T., Swann, A., Yu, J., Shorinwa, O., Murai, R., Kennedy III, M., & Schwager, M. (2024). Safer-splat: A control barrier function for safe navigation with online gaussian splatting maps. arXiv preprint arXiv:2409.09868

Kontakt: hannah.schieber@tum.de, hex-thesis.ortho@mh.tum.de