Robotics
Graduated non-convexity (GNC)
2020ActivePublished
Key
innovation
Outlier-robust estimation with a high outlier ratio, without initialization or prior detection, via a convex→non-convex homotopy.
Category
Robotics
Abstraction level
Pattern
Operation level
Robot controlTraining
How it works
A continuation parameter is introduced that initially makes the problem (nearly) convex, then is gradually changed so the cost approaches the original robust estimator. At each stage the problem is solved via iteratively reweighted least squares, updating observation weights; outliers are progressively switched off.
Problem solved
Robust cost functions are non-convex and sensitive to initialization; with many outliers (e.g. wrong loop closures, bad feature matches) classical optimization gets stuck in poor minima.