folding-state-est-gps

GPSStateEstModel — graph-based (GNN + Transformer) state estimation for variable-vertex cloth meshes. Reconstructs the full cloth mesh state from a partial point cloud observation via DDPM diffusion, conditioned on each cloth's own rest state and topology (no global template).

• Task data: folding (fold_meshes_seed_1397.h5)
• Formulation: DDPM diffusion
• Rest-position centering: pcd

Rest-position centering: pcd (legacy)

This checkpoint was trained with the pcd rest-position centering convention: the per-cloth rest (template) positions fed to the model are centered by subtracting the centroid of the observed point cloud for the same frame.

Implications:

• At inference, the rest positions must be preprocessed the same way (subtract the PCD centroid) or predictions will be wrong. Pipelines in the training repo read model.config.rest_pos_centering and do this automatically (checkpoints without the field default to pcd).
• Because the template frame is coupled to the observation, this model is sim-only in practice: with real camera point clouds the canonical template lives in a different coordinate frame, and PCD-centroid centering produces template inputs never seen in training. For real-world use, prefer a checkpoint trained with rest_pos_centering: self (e.g. fold-unfold-lift-state-est-gps-flow-v2 or fold-unfold-lift-state-est-gps-sequential).