Supplementary visualizations

Lightweight proactive anomaly detection for mobile robots

A small network watches a sliding window of onboard sensors, learns what normal driving looks like, and raises an alert when reconstruction, short-horizon prediction, and traversability disagree with that normal. The page collects the figures used for talks and for the paper supplement.

0.899

AUROC

0.828

Best F1

1.4%

False positive rate

1.20 ms

CPU, one window

Held-out comparison

Gazebo hospital corpus, cycle-held-out test, seed 42, 11,793 windows. F1 is the best F1 on that test set. The model is trained only on normal driving.

Scalar ranking and detection on the same held-out split.
Method	AUROC	AUPRC	Best F1
Isolation Forest	0.762	0.717	0.624
One-Class SVM	0.868	0.858	0.797
LSTM autoencoder	0.872	0.859	0.809
USAD	0.875	0.856	0.797
LPAD+	0.899	0.890	0.828

Detection delay in the manuscript is counted after fault onset, at a threshold frozen from normal validation. The timing number is one window on CPU. It is not an onboard Jetson measurement.

Figures

Select a group, then open a figure for the full image. Shaded fault bands in the slip plot are labels, not detector decisions.