Authors: Albert J. Zhai, Yuan Shen, Emily Y. Chen, Gloria X. Wang, Xinlei Wang, Sheng Wang, Kaiyu Guan, Shenlong Wang

Published on: April 05, 2024

Impact Score: 7.4

Arxiv code: Arxiv:2404.04242

Summary

• What is new: A novel approach using large language models for proposing candidate materials and a zero-shot kernel regression for estimating physical properties from images.
• Why this is important: Identifying materials and estimating their physical properties from visual appearance has been challenging for computers.
• What the research proposes: A method that combines large language models with language-embedded point clouds and zero-shot kernel regression for dense prediction of objects’ physical properties.
• Results: Successful application in various physical property reasoning tasks, showing accuracy in estimating mass, friction, and hardness of objects.

Technical Details

Technological frameworks used: Language-embedded point clouds

Models used: Large language models, Zero-shot kernel regression

Data used: Collection of images

Potential Impact

E-commerce, Robotics, Augmented Reality, Quality Control Industries

Want to implement this idea in a business?

We have generated a startup concept here: VisioPhysics.