Abstract
In autonomous robotic liquid transport or pesticides spraying drones, a critical destabilizing aspect exists due to unpredictable hydrodynamic sloshing force. Traditional direct measurement requires multi-axis force sensors, while conventional sensorless estimation fails to capture chaotic internal fluid motion. This research proposes a novel, hardware-free virtual sensor utilizing a parallel-branch Convolution Neural Network - Gated Recurrent Unit (CNN-GRU) deep learning architecture. By fusing fast external kinematic data from an Inertial Measurement Unit (IMU) with slow internal free-surface observations from a cheap laser and ultrasonic rangefinder, the decoupled network maps complex spatial and historical temporal dependencies without feature interference. Evaluated against traditional machine learning baseline models, the proposed CNN-GRU achieved the highest predictive accuracy offline. Furthermore, the model was successfully deployed on an edge computing for real-time hardware validation, achieving inference latencies and robust dynamic estimation (R2 = 0.85, RMSE = 0.34 N). The results confirm that physically informed feature separation and recurrent temporal modeling provide an effective, deployable solution for low cost sensorless hydrodynamic force estimation.