Terrain Classification Enhanced with Uncertainty for Space Exploration Robots from Proprioceptive Data

TL;DR

This paper tackles terrain classification for space exploration robots using proprioceptive data to avoid vulnerabilities of vision under variable conditions. It introduces uncertainty-aware time-series neural networks that incorporate MC Dropout, DropConnect, and Flipout, with hyperparameters optimized via Bayesian Optimization with Hyperband (BOHB). The study systematically compares UQ-enabled and conventional networks, reporting that MC Dropout improves calibration (lower ECE) and that IMU-only inputs and sliding-window sequence generation yield strong performance. Overall, the work demonstrates that incorporating uncertainty quantification yields trustworthy terrain classifications, supporting safer autonomous navigation in planetary missions and offering practical guidance on sensor inputs and sequence design.

Abstract

Terrain Classification is an essential task in space exploration, where unpredictable environments are difficult to observe using only exteroceptive sensors such as vision. Implementing Neural Network classifiers can have high performance but can be deemed untrustworthy as they lack transparency, which makes them unreliable for taking high-stakes decisions during mission planning. We address this by proposing Neural Networks with Uncertainty Quantification in Terrain Classification. We enable our Neural Networks with Monte Carlo Dropout, DropConnect, and Flipout in time series-capable architectures using only proprioceptive data as input. We use Bayesian Optimization with Hyperband for efficient hyperparameter optimization to find optimal models for trustworthy terrain classification.

Figures (4)

• Figure 1: The AsguardIV is a hybrid leg-wheel robot designed at the DFKI to allow navigation in unstructured environments. Its rimless wheels are simpler, more energy-efficient, and more reliable than articulated legs, adapting effectively to obstacles and uneven terrain.
• Figure 2: The experimental sites for recording data include terrain that is mostly comprised of sand and rock. Rock is represented as uneven unconsolidated rock and flat rock plateaus landscapes or inclines shown as $Class_{0}$ in the first three images. Sand is represented present in a loose and compact form shown as $Class_{1}$ in the last two images. These images serve as visual aids for labeling and are not used to train our models.
• Figure 3: All candidates, both BOHB and TSC. (a) Performance behavior is observed by comparing $F1_{cl_{0}}$ vs. $F1_{cl_{1}}$. (b) ECE scores by UQ and architecture. (c) Predictive entropy scores by UQ and architecture.
• Figure 4: Predictive accuracy versus entropy from the selected candidates for each UQ method.