Introduction
Planetary rovers like Curiosity and Perseverance primarily rely on visual systems and motor telemetry for navigation. While effective for mapping, these "eyes" cannot detect immediate subsurface hazards, such as rocks buried just beneath the dust or the subtle mechanical properties of soft sand. This lack of a real-time "sense of touch" increases the risk of wheel damage or immobilization.
To address this, researchers have developed a "tactile wheel" prototype that uses a grid of pressure sensors to provide rovers with an in-situ understanding of wheel-terrain interaction. By applying TOPO-GCN, a novel deep learning architecture combining geometric and topological data analysis, the system can identify terrain types and detect buried obstacles with significantly higher accuracy than traditional methods.
Methodology: TOPO-GCN
The proposed Topological Graph Convolutional Network (TOPO-GCN) is a hybrid architecture designed to process complex tactile data. The approach integrates three core components:
- Tactile Sensing: The rover wheel is outfitted with a 2D grid of 1,920 pressure sensors (taxels). As the wheel rotates, it generates conductivity maps representing the ground's pressure signature.
- Topological Data Analysis (TDA): To capture the structural "shape" of the terrain—such as the specific indentation patterns of a rock—the model employs Persistent Homology (PH). This technique tracks how topological features (e.g., holes or clusters) evolve across different pressure thresholds, producing Persistence Images (PIs) that are invariant to noise and minor deformations.
- Geometric Deep Learning (GDL): Unlike standard Convolutional Neural Networks (CNNs) that treat data as flat Euclidean grids, TOPO-GCN uses a Graph Convolutional Network (GCN) with a Generalized Sigma-based filter. This allows the model to better account for the intrinsic geometric structure of the wheel-terrain interface.
The final architecture fuses features from raw pressure images (processed via a standard DNN) and Persistence Images (processed via the GCN) to perform terrain classification.
Dataset and Experimental Setup
The study utilizes a unique tactile dataset generated from 163 experiments with the wheel prototype. The dataset comprises over 90,000 pressure "thumbnail" images (20x96 arrays) measured in Digital Number (DN) units. Labels include binary classifications (rock vs. no rock) across seven material types and various configurations, including surface-level and buried rocks.
Key Results
TOPO-GCN was benchmarked against several models, including standard CNNs (VGG-16) and Stochastic Gradient Boosting (SGB) models.
- Overall Accuracy: TOPO-GCN achieved 97.1%, outperforming SGB (95.3%) and VGG-16 (94.0%).
- Buried Rock Detection: The most significant breakthrough was in detecting "below-surface" rocks. TOPO-GCN achieved 31.0% accuracy, a nearly five-fold improvement over the SGB baseline (6.7%). While 31% indicates the difficulty of the task, it represents a major step forward in detecting hazards invisible to cameras.
- Computational Efficiency: Despite its complexity, TOPO-GCN remains competitive, with training times (21s per epoch) comparable to VGG-16 (27s per epoch) on modern hardware.
Discussion and Future Work
While promising, deploying these deep learning models on space-grade processors like the RAD750 remains a challenge due to strict power and compute constraints. Future research will focus on "multiple object identification" in cluttered environments and sensor fusion—combining tactile data with other instruments, such as spectrometers, to further enhance a rover's environmental awareness. By "deepening" the sense of touch, future missions can navigate more safely through the treacherous terrains of Mars and beyond.
Summary updated