Phys-Liquid: A Physics-Informed Dataset for Estimating 3D Geometry and Volume of Transparent Deformable Liquids

TL;DR

Phys-Liquid introduces a physics-informed liquid dataset and a multi-stage reconstruction pipeline to estimate 3D geometry and volume of transparent deformable liquids under container-induced deformations. The dataset uses Navier-Stokes-based simulation (via Mantaflow in Blender) to produce 97,200 images and 8,100 annotated liquid meshes across diverse scenes, colors, and rotations, complemented by six orthographic views and real-world scaling. A four-stage pipeline—segmentation with SAM2/YOLO-world, diffusion-model-based multi-view mask generation, triplane-based 3D reconstruction, and ViT-based mesh scaling—achieves superior accuracy and temporal stability compared with baselines, with demonstrated generalization to real-world data. The work enables robust perception for autonomous liquid handling in labs and provides a foundation for physics-aware perception and multi-modal liquid reasoning.

Abstract

Estimating the geometric and volumetric properties of transparent deformable liquids is challenging due to optical complexities and dynamic surface deformations induced by container movements. Autonomous robots performing precise liquid manipulation tasks, such as dispensing, aspiration, and mixing, must handle containers in ways that inevitably induce these deformations, complicating accurate liquid state assessment. Current datasets lack comprehensive physics-informed simulation data representing realistic liquid behaviors under diverse dynamic scenarios. To bridge this gap, we introduce Phys-Liquid, a physics-informed dataset comprising 97,200 simulation images and corresponding 3D meshes, capturing liquid dynamics across multiple laboratory scenes, lighting conditions, liquid colors, and container rotations. To validate the realism and effectiveness of Phys-Liquid, we propose a four-stage reconstruction and estimation pipeline involving liquid segmentation, multi-view mask generation, 3D mesh reconstruction, and real-world scaling. Experimental results demonstrate improved accuracy and consistency in reconstructing liquid geometry and volume, outperforming existing benchmarks. The dataset and associated validation methods facilitate future advancements in transparent liquid perception tasks. The dataset and code are available at https://dualtransparency.github.io/Phys-Liquid/.

Figures (19)

• Figure 1: Simulation samples of five transparent containers under five distinct simulation sets from Phys-Liquid, showing variations in laboratory scenes, lighting conditions, container rotations, liquid colors, volumes, and deformations.
• Figure 2: The setting of six orthographic camera views for the representation of triplane in the scene setting.
• Figure 3: The liquid mesh formed by particles.
• Figure 4: Multi-view and temporal representations.
• Figure 5: Validation of simulation realism by comparing liquid deformations with real-world experiments.