Classification of liquid viscosity in containers based on force sensing

Humans tend to perceive the properties of an object by the way they interact with the objects around them, for example, by simply shaking a container to determine the viscosity of the liquid inside it. Nowadays, as robots are widely deployed in life, we want to give them the same perception capabilities to make subsequent correct motion decisions based on the perception results. Existing studies on liquid viscosity perception are mainly based on visual sensors, but visual perception is hindered when the liquid is obscured by an opaque or closed container. To make up for the shortage of visual perception, we innovatively propose a six-axis force/torque sensors perception method based on the end of the robot arm. By collecting the oscillation information of the liquid in the container after shaking, using the CATCH22 time series analysis package to extract important features from the time series data, and combining with Neighborhood Component Feature (NCA) technology for model optimization, we tested the classification effect of the model on seven liquids in the viscosity range of 1-1000cst and common liquids in daily life respectively. The experimental results show that our method achieves 99.6% accuracy in classifying water, milk, yogurt, and detergent, which are common liquids in life, and 93.3% accuracy on seven liquids with similar viscosity, proving the feasibility of the force-information-based approach to perceiving liquid viscosity.