Unsupervised Anomaly Detection for Autonomous Robots via Mahalanobis SVDD with Audio-IMU Fusion

TL;DR

This work explores the use of audio and inertial measurement unit (IMU) sensors to detect underlying anomalies in autonomous mobile robots, such as collisions and internal mechanical faults, and proposes an unsupervised anomaly detection framework based on Mahalanobis Support Vector Data Description (M-SVDD).

Abstract

Reliable anomaly detection is essential for ensuring the safety of autonomous robots, particularly when conventional detection systems based on vision or LiDAR become unreliable in adverse or unpredictable conditions. In such scenarios, alternative sensing modalities are needed to provide timely and robust feedback. To this end, we explore the use of audio and inertial measurement unit (IMU) sensors to detect underlying anomalies in autonomous mobile robots, such as collisions and internal mechanical faults. Furthermore, to address the challenge of limited labeled anomaly data, we propose an unsupervised anomaly detection framework based on Mahalanobis Support Vector Data Description (M-SVDD). In contrast to conventional SVDD methods that rely on Euclidean distance and assume isotropic feature distributions, our approach employs the Mahalanobis distance to adaptively scale feature dimensions and capture inter-feature correlations, enabling more expressive decision boundaries. In addition, a reconstruction-based auxiliary branch is introduced to preserve feature diversity and prevent representation collapse, further enhancing the robustness of anomaly detection. Extensive experiments on a collected mobile robot dataset and four public datasets demonstrate the effectiveness of the proposed method, as shown in the video https://youtu.be/yh1tn6DDD4A. Code and dataset are available at https://github.com/jamesyang7/M-SVDD.

Figures (11)

• Figure 1: Our proposed method integrates audio and IMU data for anomaly detection in autonomous robots. The network is trained exclusively on normal operational data and effectively identifies anomalies during inference.
• Figure 2: The overall workflow of the proposed unsupervised anomaly detection framework. In the training stage, audio and IMU data collected under normal conditions are input into the corresponding encoders for feature extraction. A cross-attention module is then used to fuse the features, which are fed into the M-SVDD branch to learn an ellipsoidal boundary enclosing normal data. Meanwhile, a reconstruction branch is applied to recover the original inputs, preventing feature collapse and improving detection accuracy. During inference, anomalies are detected based on reconstruction error and the Mahalanobis distance between the embedded feature and the center of the learned boundary.
• Figure 3: The cross-attention module for audio-IMU feature fusion.
• Figure 4: The experimental platform utilized for data collection.
• Figure 5: Examples of selected samples from normal driving and anomalous signals.