Diagnostic Runtime Monitoring with Martingales

TL;DR

The paper tackles robustness of learning-enabled robotics to distribution shifts by introducing online, multi-martingale runtime monitors that diagnose not just the occurrence but the type of shift. By deploying separate monitors over inputs, intermediate representations, and outputs, the approach yields fast shift detection with controlled false alarms and enables targeted interventions tailored to the identified cause. Empirical results on photorealistic X-Plane simulations and hardware free-flyer setups show substantially faster Shift detection (up to several-fold) and improved lifecycle performance when using type-specific interventions versus generic maintenance. This framework supports safer, more reliable deployment of learning-based robotics by enabling rapid, actionable diagnostics and mitigations.

Abstract

Machine learning systems deployed in safety-critical robotics settings must be robust to distribution shifts. However, system designers must understand the cause of a distribution shift in order to implement the appropriate intervention or mitigation strategy and prevent system failure. In this paper, we present a novel framework for diagnosing distribution shifts in a streaming fashion by deploying multiple stochastic martingales simultaneously. We show that knowledge of the underlying cause of a distribution shift can lead to proper interventions over the lifecycle of a deployed system. Our experimental framework can easily be adapted to different types of distribution shifts, models, and datasets. We find that our method outperforms existing work on diagnosing distribution shifts in terms of speed, accuracy, and flexibility, and validate the efficiency of our model in both simulated and live hardware settings.

Figures (10)

• Figure 1: Overview of our high-level approach. Learning-enabled robotics systems are trained on data from a finite set of environments. When deployed, these systems may operate in distribution-shifted conditions, resulting in erroneous predictions. Our method issues an alert if conditions change, and alerts users of a probable underlying cause. Knowledge of the underlying cause informs the choice of the proper intervention method to restore system performance.
• Figure 2: Images generated from the X-Plane 11 flight simulator, with (\ref{['fig:camera_correct']}) a standard camera angle, (\ref{['fig:camera_perturbed']}) a shifted environment, and (\ref{['fig:xplane_camera_results']}) sensor degradation. We define separate martingales to identify each type of shift.
• Figure 3: Martingale values for our method and the CM method, in the presence of a sensor degradation shift, which starts occurring at $t=0$. The martingales grow as they detect the shift and an alert is issued when the martingale value exceeds the threshold of 100. Our method raises an alert much sooner on average (14.92 iterations) compared to the CM method (38.10 iterations), showing that our method detects this distribution shift faster than existing methods.
• Figure 4: Martingale values for a brightness shift when multiple martingales are deployed simultaneously. All three martingales detect a shift, but the brightness shift martingale issues a warning signal before the other two martingales.
• Figure 5: Free-flyer robot platform

Theorems & Definitions (2)

• definition thmcounterdefinition: Martingale
• Proposition 1: Doob's Inequality