Event

Dense-Armor Open-Sources SONAR, an SGD-Based One-Class Joint Anomaly Detection Module

1 reports · 1 sources

The Dense-Armor project has open-sourced the SONAR module, which uses random Fourier features to approximate an RBF kernel and adds strong convex regularization, allowing the one-class boundary to be updated online per sample via SGD, for determining whether robot joint residual readings deviate from a healthy state: each reading receives a score, where a negative value indicates inside the boundary (normal) and a positive value indicates outside the boundary (anomaly), with larger values indicating farther deviation. Under the benchmark default parameters (n_features_rff=128, n_init=20), when processing the arm data stream on a single CPU thread, learn_one plus score_one takes 0.131 milliseconds per sample. The default threshold None uses the 0.95 quantile of the most recent 500 scores; when fewer than 50 scores have been seen, the effective threshold is 1e9, and no samples will be flagged. The arm data stream has only 100 healthy samples, and after subtracting n_init=20 there are only 80 SGD steps, far fewer than the thousands of steps in the paper's experiments, with an ROC-AUC of 0.856. The paper additionally proposes the SONARC variant for non-stationary data streams (with change point detection and restart), which this module does not implement.