Explainable AI-Enhanced Human Activity Recognition for Human–Robot Collaboration in Agriculture

Abstract

This study addresses a critical gap in human activity recognition (HAR) research by enhancing both the explainability and efficiency of activity classification in collaborative human–robot systems, particularly in agricultural environments. While traditional HAR models often prioritize improving overall classification accuracy, they typically lack transparency in how sensor data contribute to decision-making. To fill this gap, this study integrates explainable artificial intelligence, specifically SHapley Additive exPlanations (SHAP), thus enhancing the interpretability of the model. Data were collected from 20 participants who wore five inertial measurement units (IMUs) at various body positions while performing material handling tasks involving an unmanned ground vehicle in a field collaborative harvesting scenario. The results highlight the central role of torso-mounted sensors, particularly in the lumbar region, cervix, and chest, in capturing core movements, while wrist sensors provided useful complementary information, especially for load-related activities. The XGBoost-based model, selected mainly for allowing an in-depth analysis of feature contributions by considerably reducing the complexity of calculations, demonstrated strong performance in HAR. The findings indicate that future research should focus on enlarging the dataset, investigating the use of additional sensors and sensor placements, and performing real-world trials to enhance the model’s generalizability and adaptability for practical agricultural applications.

Keywords

• XGBoost classification
• SHapley Additive exPlanations (SHAP)
• feature importance
• field experiment
• material handling tasks
• wearable sensors