sandeeptshelvan/physics-prior-mlp

Physics-prior MLP

These weights were trained in this project. They are the epoch-10 checkpoint from a 14-epoch run of a per-Gaussian MLP for SceneAgent section 3.4 (Hollis, Fan, and Yang, arXiv:2610.34745).

This repository is not a released third-party checkpoint.

Code and notes: sandy-t-shelvan/physics-prior-mlp.

Checkpoint

File	physics_prior_splat.npz (260290 bytes)
Split	n_fitted 3736, n_train 3274, n_validation 139, n_test 323
epochs_ran	14
checkpoint_epoch	10

The saved epoch is the earliest epoch with the lowest validation loss. The test split was scored once and did not choose the epoch. physics_prior.npz is an earlier file and is not this model.

Scores

Rounded to three decimals from train_report_splat.json. Material, body, and collision are accuracies. The other columns are mean absolute error.

Split	Loss	Material	Body	Collision	Log-mass MAE	Log-density MAE	Rigidity MAE	Deformability MAE
Train	1.350	0.953	0.993	0.831	0.292	0.494	0.035	0.033
Validation	2.293	0.928	0.957	0.820	0.383	0.577	0.051	0.047
Test	2.088	0.889	0.991	0.768	0.352	0.572	0.037	0.035

Files

• physics_prior_splat.npz — model weights
• train_report_splat.json — split sizes, losses, and the full report

Load the checkpoint with NumPy. Meshes and per-Gaussian prior_input.npy files are not in this repository.