Soil Type Recognition for Robotic Sampling in Deep Space Exploration using Haptic Information

Abstract

Sampling by drilling is a typical method to obtain lunar soil samples. To guide the decision of the sampling process, it is necessary to accurately identify the soil type at the sampler probe location with less sensory data to reduce the cost in deep space exploration. In this paper, we study the advanced learning models based on the self-attention mechanism for the task of soil type classification in deep space exploration. Our model only depends on one-dimensional force data to achieve accurate classification of soil types. The mechanical data collected from five different simulated soils were used for training. In the cases with less force sensory data, the classification results of our model outperformed the traditional learning methods and the prediction accuracy reached 100% in our test. Our model also has better performance in terms of the Macro-Precision, Macro-Recall, and Macro-F1 score metrics, showing its potential for robotic sampling in deep space exploration.