Fast certifiable estimators solve locally and then certify the candidate with cheap linear algebra, but many robotics relaxations are degenerate and break that check. CP-Cert starts from the local candidate and moves to a nearby point on the central path, where a valid certificate is easy to obtain, using indirect linear algebra, sparsity and parallelism. It runs up to three orders of magnitude faster than direct SDP solvers on simulated matrix-weighted pose registration and point-cloud data association, and is combined into a certifiable, outlier-robust pose pipeline run on real data.
Relevance
Certifiable perception has been held back by speed on degenerate problems, and outlier-robust registration is where degeneracy shows up most. See robust estimation under robust robotics.
Links
- arXiv
- Subfield: Safety and verification, Mapping and localisation