The Sound of Water: Inferring Physical Properties From Pouring Liquids

Area of Science:

• Acoustics
• Physics
• Machine Learning
• Robotics

Background:

• Liquid pouring is a common activity with complex physics.
• Understanding pouring dynamics typically requires visual or direct measurement.
• Acoustic signals offer a non-invasive way to infer physical properties.

Purpose of the Study:

• To automatically infer physical properties of liquid pouring using only sound.
• To investigate the relationship between audio observations and liquid pouring physics.
• To develop a model capable of estimating liquid level, container characteristics, and pouring dynamics.

Main Methods:

• Theoretical analysis demonstrating pitch (fundamental frequency) relates to pouring properties.
• Training a pitch detection model using simulated and physics-informed visual data.
• Creation and utilization of a novel, large dataset of real-world pouring videos.
• Validation of the model's inference capabilities on real-world pouring data.

Main Results:

• The fundamental frequency of pouring sounds theoretically correlates with key physical properties.
• A trained pitch detection model successfully infers liquid level, container shape/size, and pouring rate from audio.
• The model demonstrates robust generalization across diverse container types, datasets, and unconstrained YouTube videos.
• Significant accuracy in estimating pouring parameters was achieved using only acoustic data.

Conclusions:

• Sound analysis provides a viable method for inferring liquid pouring physics.
• This research bridges acoustics, physics, and machine learning for practical applications.
• The developed approach enhances multisensory perception, particularly for robotic pouring tasks.