Safety Enhancement in Planetary Rovers: Early Detection of Tip-over Risks Using Autoencoders

TL;DR

This work addresses the safety challenge of tip-over risks for the rimless-wheeled AsguardIV planetary rover by forecasting IMU sequences with a Long Short-Term Memory (LSTM) Autoencoder and detecting anomalies indicative of incipient tip-overs. The methodology frames the problem as a sequence-to-sequence forecasting task, using six-feature IMU data and an unsupervised training regime to learn normal operating dynamics. Through Bayesian optimization with Hyperband (BOHB), the study identifies a lightweight, single-layer LSTM that achieves high forecast accuracy, with an impressive $R^{2}=0.969$ on test data. Tip-over detection relies on thresholds derived from training loss distributions for both MSE and MAE, enabling near-simultaneous multi-axis risk signaling across sensor channels. The work demonstrates the practical potential of autoencoder-based predictive safety for autonomous rover missions in challenging terrains, with a concrete horizon of a 250 ms input and 50 ms output for effective early warning.

Abstract

Autonomous robots consistently encounter unforeseen dangerous situations during exploration missions. The characteristic rimless wheels in the AsguardIV rover allow it to overcome challenging terrains. However, steep slopes or difficult maneuvers can cause the rover to tip over and threaten the completion of a mission. This work focuses on identifying early signs or initial stages for potential tip-over events to predict and detect these critical moments before they fully occur, possibly preventing accidents and enhancing the safety and stability of the rover during its exploration mission. Inertial Measurement Units (IMU) readings are used to develop compact, robust, and efficient Autoencoders that combine the power of sequence processing of Long Short-Term Memory Networks (LSTM). By leveraging LSTM-based Autoencoders, this work contributes predictive capabilities for detecting tip-over risks and developing safety measures for more reliable exploration missions.

Figures (8)

• Figure 1: The AsguardIV rover shown traversing an inclined lunar analog plane during data collection.
• Figure 2: Architecture of the LSTM Autoencoder. Using LSTM layers enables the Autoencoder to learn temporal structures to forecast sequences from reconstructed inputs mythesis.
• Figure 3: Examples of the AsguardIV performing two navigation executions mythesis. IMU plots across all axes. Figure (a) illustrates a mobility risk marked in red. Figure (b) shows a navigation sequence with no mobility risks, however, it marks in green a sequence that could be mistaken as one.
• Figure 4: First test results showing $R^{2}$ scores per different Autoencoder layer depths per non-BOHB setting.
• Figure 5: Tests trials showing the forecast signals (red) from the in put sequence (green). It can be observed that the normal IMU signals are accurately predicted whereas the anomalies are not, enabling effective detection of potential tip-over risks.