https://louenpottier.github.io/lagsplat.html
LaGSplat infers the physics of an object from a few seconds of video: a visual encoder projects each frame into a low-dimensional latent space, a Lagrangian Neural Network learns the equations of motion there, then a Gaussian Splatting decoder reconstructs the scene in real time. An external force can be applied at any instant, anywhere in the scene, even though the training data contains no measured force.
LaGSplat infers the physics of an object from a few seconds of video: a visual encoder projects each frame into a low-dimensional latent space, a Lagrangian Neural Network learns the equations of motion there, then a Gaussian Splatting decoder reconstructs the scene in real time. An external force can be applied at any instant, anywhere in the scene, even though the training data contains no measured force.