Semantic Segmentation for Robot Localization
What
Most production-ready robot localization pipelines rely on geometric features and known maps. Semantic information — what an object is, not just where edges are — is usually treated as a downstream task or an aid for loop closure rather than a first-class signal in the localization estimate.
This work explores localization methods that put semantic segmentation and object registration at the core of the pose estimate, instead of bolting them on after a geometric front end.
Why
The application target is mobile robots operating in environments where pure geometric SLAM degrades (repetitive structure, dynamic scenes, sparse texture) but where semantic content is rich and stable. The hypothesis is that semantic-first formulations can extract more reliable pose information from these scenes than geometry-first methods that treat semantics as an afterthought.
What’s coming
The full writeup will cover:
- Problem framing and where current semantic-aware SLAM falls short
- The proposed method and its formulation
- Experimental setup and the datasets I’m evaluating on
- Ablations and what I think the load-bearing components are
- Limitations and where it doesn’t work
- What’s next
Until then, you can find me on the Mines campus or via the contact links below.