Leveraging Energy Features for Surface Classification with Deep Learning: A Comparative Analysis Across Three Independent Datasets

A new study finds that energy-derived features can classify surfaces for mobile robots with 85-90% accuracy on their own, and boost inertial-sensor accuracy by 1-2% when combined, outperforming earlier benchmarks across three public datasets. The work, posted to arXiv on 17 June 2026, tested recurrent neural networks, convolutional neural networks, encoder-only transformers, and Mamba state-space models under automated hyperparameter tuning and input-sequence-length optimization [1]. Convolutional neural networks delivered the highest overall performance, while classifiers using only energy features reached 85-90% accuracy — roughly 5-10 percentage points below the 96-99% achieved when energy and inertial data were fused [1]. Augmenting inertial data with energy features yielded a consistent mean accuracy improvement of 1-2% [1]. The authors conclude that energy-only classifiers are accurate enough for standalone deployment and provide a reliable gain when paired with other sensing modalities [1]. Surface classification is a well-studied problem in mobile robotics, but energy-based approaches have received comparatively little attention despite promising results in constrained environments [2]. The new study addresses that gap by evaluating performance across three independent, publicly available datasets, making it one of the most comprehensive comparisons of deep-learning architectures for this task [2][3]. Other recent research has explored surface and material classification using different sensor modalities. A LiDAR-driven machine-learning framework classified indoor surfaces into semi-specular and low-specular categories for millimeter-wave and sub-terahertz network planning, with Random Forest models providing the best accuracy-robustness trade-off [5]. Separately, a multimodal system combining vision and tactile data achieved 99.4% accuracy with a Surformer v1 model, while a Multimodal CNN reached 100% accuracy but required 48.3 million parameters and 5.07 ms of inference time per sample — far heavier than the 673,321-parameter Surformer v1 at 0.73 ms [4]. Those findings underscore a broader trend: lightweight models with carefully engineered features can approach the accuracy of much larger systems while remaining practical for real-time robotic applications where computational resources are limited [4][5]. The energy-feature study extends this principle to inertial and energy-based sensing, showing that even a simple feature set can support high-accuracy surface classification without the hardware burden of vision or tactile arrays [1][2].