Improving Methods for Multi-Terrain Classification Beyond Visual Perception
Published in 2021 Fifth IEEE International Conference on Robotic Computing (IRC), 2022
Terrain classification in mixed-surface unstructured environments is key for safe navigation, energy efficiency, and anticipating motion volatility. This is particularly true for dynamically moving legged platforms which are highly impacted by foot ground interactions. This research demonstrates terrain classification using a long short-term memory (LSTM) model trained on actuator time series data, particularly the difference in center-of-pressure (COP) and leg forces. The LSTM COPForce model showed a 97.5% accuracy in classification on three outdoor surfaces with small amounts of data and no additional sensors.
Recommended citation:
C. Allred, M. Russell, M. Harper and J. Pusey, “Improving Methods for Multi-Terrain Classification Beyond Visual Perception,” 2021 Fifth IEEE International Conference on Robotic Computing (IRC), Taichung, Taiwan, 2021, pp. 96-99, doi: 10.1109/IRC52146.2021.00022.