Integrating Topological Object Recognition into Semantic SLAM for Unseen Cluttered Environments

Avania Bhattacharya, Ekta Umesh Samani, Ashis Gopal Banerjee

Keywords: Semantic SLAM, object recognition, topological descriptors

Abstract: Reliable semantic mapping is challenging for mobile robots in previously unseen cluttered environments, as vision-based SLAM pipelines are often sensitive to occlusion, viewpoint variation, and environmental clutter. We integrate THOR2, a topological object recognition framework, into Kimera Semantics for object-level mapping and show that our approach achieves higher recognition accuracy than Mask R-CNN, YOLOv8, and DINOv2 RGB/RGB-D baselines across clutter levels and robot trajectories. Unlike conventional deep learning approaches, our method leverages domain-invariant shape-based features and improves robustness to partial observations. Integrating topological object recognition into semantic mapping is therefore a meaningful step toward scene understanding that generalizes to previously unseen cluttered environments and better supports downstream robot autonomy.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 30