Learning-Based Fault Detection for Legged Robots in Remote Dynamic Environments

1. Problem Statement

Legged robots operating in hazardous, remote environments (e.g., disaster response, radiation monitoring) face significant risks of physical damage. Unlike humans or animals, quadruped robots cannot naturally instinctively adjust their locomotion to compensate for a severely debilitated limb.

• The Core Challenge: Existing fault recovery strategies (such as tripedal gaits) require the robot to first identify and locate the specific damaged limb.
• Limitations of Current Methods:
  • Model-based approaches require highly accurate dynamic models of the robot, which are often unavailable or difficult to derive.
  • Evolutionary algorithms for anticipated faults are computationally expensive and time-consuming.
  • Supervised learning requires labeled datasets, which are difficult to acquire in real-time, unstructured environments where the robot is actively collecting data without prior knowledge of faults.
• Goal: Develop a computationally efficient, real-time capable method to detect and locate a single limb failure using proprioceptive sensor data, specifically utilizing unlabeled data (unsupervised learning) to identify anomalies.

2. Methodology

The authors propose an offline unsupervised deep learning approach using an Autoencoder architecture to perform anomaly detection.

A. Dataset Construction

• Source: Data was collected from a physical quadruped robot (Minitaur) performing "fore-aft pronking" gaits.
• Conditions: The dataset includes 170 trials across five conditions:
  1. None (N): Intact robot (Normal).
  2. LF: Left-Front limb damaged.
  3. RF: Right-Front limb damaged.
  4. LB: Left-Back limb damaged.
  5. RB: Right-Back limb damaged.
• Features: Each data point includes proprioceptive sensor measurements sampled at 60 Hz:
  • Limb position and velocity (body coordinates).
  • Hip motor currents ($\alpha$ and $\beta$).
• Preprocessing: Data was filtered to include only moving (pronking) phases and scaled using MinMaxScaler (range 0–1).

B. Model Architecture: Autoencoder

The system uses an Autoencoder to learn the "normal" distribution of gait data.

• Training Strategy: The model is trained exclusively on the "None" (intact) dataset. It learns to compress (encode) and reconstruct normal gait patterns.
• Architecture:
  • Encoder: 3 layers (8 neurons $\to$ 4 neurons $\to$ 2 neurons) using ReLU activation. This creates a bottleneck (low-dimensional representation).
  • Decoder: 3 layers (4 neurons $\to$ 8 neurons $\to$ 4 neurons) using ReLU activation.
• Loss Function: Mean Squared Error (MSE) between the input and the reconstructed output.

C. Anomaly Detection Logic

• Principle: The model minimizes reconstruction error for normal data. When presented with damaged limb data (an anomaly), the model fails to reconstruct it accurately, resulting in a high reconstruction loss.
• Thresholding: A threshold is calculated based on the training data's reconstruction loss statistics:
  $$\text{Threshold} = \mu_{\text{training}} + x \cdot \sigma_{\text{training}}$$
  Where $\mu$ is the mean and $\sigma$ is the standard deviation of the training loss. The multiplier $x$ is tuned to prioritize minimizing false negatives (missing a fault) over false positives.
• Decision: If the reconstruction loss of a new sample exceeds the threshold, the sample is flagged as a fault.

3. Key Contributions

1. Unsupervised Fault Detection: Demonstrated the viability of using autoencoders to detect limb faults without requiring labeled fault data, making the approach suitable for real-world deployment where fault labels are unknown.
2. Physical Implementation: Validated the algorithm using empirical data collected from a physical robot, moving beyond purely simulation-based studies.
3. Integration Framework: Proposed a workflow where the offline learning module outputs a specific limb fault diagnosis to a low-level controller, enabling the robot to switch to a specific tripedal gait (e.g., "fore-aft tripedal pronking") appropriate for the missing limb.
4. Specific Limb Localization: The system successfully distinguishes not just that a fault exists, but which limb is damaged (e.g., distinguishing LF from RF damage).

4. Experimental Results

The model was evaluated using a 60:20:20 split (Train:Validation:Test) on the intact data, with damaged data used solely for testing.

• Reconstruction Performance:
  • Intact Data: The model achieved low reconstruction loss, indicating high accuracy in learning normal gait patterns.
  • Damaged Data: The model produced significantly higher reconstruction losses for damaged limbs, clearly separating them from normal data distributions.
• Detection Accuracy:
  • False Negatives (Missed Faults): Extremely low. The model correctly identified 99.89% of Left-Front (LF) damaged samples and 96.86% of Right-Front (RF) damaged samples as anomalies.
  • False Positives (False Alarms): The model correctly identified 84.39% of intact samples as non-damaged (LF model) and 82.48% (RF model).
• Visualization: Histograms of reconstruction loss showed distinct separation between the distributions of normal and damaged data, with the threshold effectively capturing the anomaly boundary.

5. Significance and Future Work

• Significance: This work bridges the gap between theoretical fault recovery gaits and practical deployment. By enabling autonomous fault detection, robots can survive and complete missions in environments where human intervention is impossible. The unsupervised nature of the solution makes it robust to the unpredictability of real-world damage.
• Future Directions:
  • Ensemble Modeling: Investigating whether combining individual limb-detection models into a single ensemble model improves overall performance.
  • Closed-Loop Integration: Integrating this detection module with the previously developed bio-inspired tripedal gait controllers to create a fully autonomous fault-recovery system.
  • Dynamic Environments: Testing the system in highly complex, dynamic settings beyond controlled laboratory conditions.

In conclusion, the paper successfully demonstrates that an unsupervised autoencoder trained on proprioceptive data can effectively detect and localize single-limb failures in quadruped robots, providing a critical step toward resilient autonomous field operations.