CoSMO turns the VLM embeddings in a semantic map into conformal coverage sets, then inflates obstacles by class so each hazard gets the clearance it needs. The hazard vocabulary can grow during exploration and the data can drift, two things existing conformal methods assume away. On Matterport3D it cuts semantic collisions by 61.6% against a geometry-only baseline, and on a quadruped it holds 90.73% long-run coverage against a 90% target.
Relevance
VLM confidence scores are poorly calibrated, and this is one concrete way to wrap them in a guarantee that survives an open vocabulary. Relevant to robust robotics and to open-vocabulary mapping, where few systems say how wrong a label is allowed to be.
Links
- arXiv · project page
- Subfield: Planning and navigation, Semantic mapping, Perception