Enhancing Rover Mobility Monitoring: Autoencoder-driven Anomaly Detection for Curiosity

TL;DR

This work addresses the challenge of automatic drive anomaly detection for the Curiosity rover by developing CAIDDA, two undercomplete autoencoder variants that operate on rich mobility telemetry. Anomaly scores are derived from the reconstruction error $\mathbf{e}=\mathbf{x}-\tilde{\mathbf{x}}$ with $a=\|\mathbf{e}\|_1$, and a $99.9$th percentile threshold is used to flag unusual drives, enabling early, subtle telemetry pattern detection that may escape human review. CAIDDA-Prime uses full feature coverage (322 features) and shows sensitivity to acceleration-related events, while CAIDDA-Refined excludes acceleration features (301 features) to better detect actuator-telemetry anomalies such as wheelies and slips. The study demonstrates the practical value of unsupervised, low-dimensional representations for proactive rover health monitoring, with implications for safer, more reliable planetary exploration via enhanced downlink screening and maintenance planning.

Abstract

Over eleven years into its mission, the Mars Science Laboratory remains vital to NASA's Mars exploration. Safeguarding the rover's long-term functionality is a top mission priority. In this study, we introduce and test undercomplete autoencoder models for detecting drive anomalies, using telemetry data from wheel actuators, the Rover Inertial Measurement Unit (RIMU), and the suspension system. Our approach enhances post-drive data analysis during tactical downlink sessions. We explore various model architectures and input features to understand their impact on performance. Evaluating the models involves testing them on unseen data to mimic real-world scenarios. Our experiments demonstrate the undercomplete autoencoder model's effectiveness in detecting drive anomalies within the Curiosity rover dataset. Remarkably, the model even identifies subtle anomalous telemetry patterns missed by human operators. Additionally, we provide insights into optimal design choices by comparing different model architectures and input features. The model's ability to capture inconspicuous anomalies, potentially indicating early-stage failures, holds promise for the field, by improving the reliability and safety of future planetary exploration missions through early anomaly detection and proactive maintenance.

Figures (4)

• Figure 1: Overview of Rover Wheel Annotation
• Figure 2: Acceleration during Rock Traversal
• Figure 3: Proposed Autoencoder Architecture
• Figure 4: Reconstruction Error a and Slip Ratio for Sol 3839