KnockGS:interaction-Grounded Calibrationof Physical Gaussian Representations
KnockGS uses an observed force response to estimate a Gaussian object’s elasticity and density scales.
The paper frames physical simulation as a calibration signal, not just a forward renderer. Its method extracts temporal response features from known-force dynamics, freezes the estimated material scales, and writes them back into the simulator. The authors test recovery against hidden ground truth and measure re-simulated behavior with particle trajectories, response statistics, and rendered-frame quality. Across five held-out material targets, they report better scale recovery than retrieval, global regression, or a fixed default, with predictions holding under different force directions and magnitudes. ArXiv · AI/CL/LG's note
The paper frames physical simulation as a calibration signal, not just a forward renderer. Its method extracts temporal response features from known-force dynamics, freezes the estimated material scales, and writes them back into the simulator. The authors test recovery against hidden ground truth and measure re-simulated behavior with particle trajectories, response statistics, and rendered-frame quality. Across five held-out material targets, they report better scale recovery than retrieval, global regression, or a fixed default, with predictions holding under different force directions and magnitudes. ArXiv · AI/CL/LG's note
score 5