Statistical Multivariate Node-Level Anomaly Localization for Robot Runtime Monitoring

As autonomous robots move from the lab into everyday environments, monitoring their software health at runtime becomes a key requirement for trustworthy operation. A running ROS 2 system exposes dozens of metric streams per node, and in practice, the question is often not whether something is wrong, but which node is responsible. Existing runtime detectors typically treat each metric stream independently, flagging outliers one series at a time and localizing the culprit only by counting how many of a node's streams are flagged. We propose a multivariate approach, based on the Mahalanobis distance, which describes a node's nominal behaviour via the mean vector and covariance matrix of its runtime metrics, capturing anomalies that arise from correlations between metrics which look normal in isolation. Each node is scored at every time window both against its own nominal profile (catching the progressive degradation of a known node, such as a CPU runaway or a memory leak) and against a pooled model of typical node behaviour, which flags newly appearing or grossly atypical nodes. We evaluate the approach on a simulated TIAGo Pro robot in Gazebo across node-level fault scenarios spanning three telemetry-observability regimes, comparing against univariate baselines. When the fault leaves a footprint in the monitored metrics, the applicable Mahalanobis score ranks the faulty node first in nearly all affected windows. The pooled score achieves lower held-out false-positive rates than the pooled univariate baselines and, on one fault class whose individual metrics remain marginally unremarkable, ranks the culprit first in 12 of 14 affected windows, whereas both univariate rankings fail; no method crosses its detection threshold on this class. Across regimes, localization quality tracks what the telemetry exposes: observability, rather than the choice of statistic, emerges as the binding constraint.