Collision-aware interactive simulation using graph neural networks.

Area of Science:

• Computer Graphics
• Physics Simulation
• Machine Learning

Background:

• Deep simulations offer high acceleration but lack effective collision handling.
• Existing methods struggle with accurate tool-object collision detection and response.

Purpose of the Study:

• To propose a deep interactive physical simulation framework for robust tool-object collision management.
• To enhance dynamic information prediction by incorporating collision states.

Main Methods:

• Utilized a graph neural network as the core model.
• Introduced a collision-aware recursive regression module using interpenetration distances from vertex-face and edge-edge tests.
• Developed a novel self-supervised collision term for compact collision response.

Main Results:

• The proposed framework effectively predicts dynamic information considering collision states.
• Significantly reduced interpenetration artifacts in simulations.
• Maintained high simulation efficiency.

Conclusions:

• The deep interactive physical simulation framework provides an effective solution for tool-object collisions.
• The method balances accuracy in collision response with computational efficiency.