Friction Prediction from Motion

Physics-informed neural networks and Blender-augmented datasets to predict surface friction

Summary

Created an augmented dataset using Blender to simulate interactions between textured spheres and surfaces, and trained physics-informed neural networks to estimate friction coefficients from video.

Highlights

  • Pipeline for high-fidelity Blender-based data augmentation and synthetic rendering to expand training diversity.
  • Combined visual object detection with PINN models to estimate friction with strong generalization to novel textures.

(No public repo available)