dexgarmentlab-folding-lifting-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: DexGarmentLab mixed-garment fold + lift-place demos, point-cloud-cleaned (dexgarmentlab_folding_lifting_meshes_clean.h5, Cloth-splatters/dexgarmentlab-folding-lifting-meshes-clean)
• Formulation: DDPM diffusion
• Rest-position centering: self
• Cross-attention mode: parallel
• Max vertices per mesh: 2048
• Best validation loss: 0.0002055970478668066 (checkpoint in model/ is checkpoint-best)
• Training run: dexgarment_ddpm_clean_2026-08-05_14-28-23_475354 (full config in config.yml)

Retrained on cleaned point clouds (2026-08-06)

This revision supersedes the 2026-08-04 upload, which was trained on dexgarmentlab_folding_lifting_meshes.h5 before the point-cloud audit. Roughly half of those observations carried scene geometry (floor / fixture returns) mixed into the cloth points; the cleaned dataset removes it and trims the affected frames. The previous revision is still reachable by commit hash.

Validation losses are not comparable across that change — the split itself changed, so a number from the pre-clean revision does not sit on the same scale as the one above.

Rest-position centering: self

This checkpoint was trained with the self rest-position centering convention: the per-cloth rest (template) positions fed to the model are centered by subtracting the rest mesh's own (valid-vertex) mean, independent of the observed point cloud.

Implications:

• The model is translation-invariant in rest space, which makes it suitable for real-world inference: the canonical template can live in any coordinate frame relative to the camera point cloud.
• At inference, the rest positions must be self-centered the same way. Pipelines in the training repo read model.config.rest_pos_centering (recorded as "self" in this checkpoint's config.json) and do this automatically.
• Not interchangeable with checkpoints trained with the legacy pcd convention (all state-estimation checkpoints in this org trained on or before 2026-05-10).

v2 (2026-08-20) — this revision

Retrained with the v2 recipe: EMA weights (shipped in model/), heavy occlusion/outlier augmentation (fold-unfold-lift set, scale 0.8-1.25, no yaw). Best validation MSE 2.72e-4 m² (previous release: 6.26e-4); held-out first-frame reconstruction better on 7/8 test cloths, and the TNLC_Jacket001_0 hard start improves from rank 136 to rank 120 in the 239-mesh identification probe. Sequential cross-attention, 200k steps. The previous release remains available at revision 204401acd818e8ef9c3ad4b014bdd34206c149a3.

v2-tf2 (2026-08-21) — this revision

The v2 recipe (EMA + heavy augmentation) plus two-frame temporal training (num_context_frames: 2): during training the model cross-attends to the current cloud and one from 1-5 frames earlier. At inference it runs single-frame (pcd_prev=None duplicates the current cloud — the trajectory-start pair it trained on); the temporal input adds 2 % if supplied. The two-frame training acts as a temporal-consistency regularizer and yields the best checkpoint on every measured axis: best val (2.38e-4 m²), frame-sweep worst observation 20 of 239 (predecessors: 56-107), the TNLC_Jacket001_0 hard start at identification rank 3 (was 120-136), held-out identify-then-track kept 7/8 with MAP correct 6/8. Samplers saturate at 3 denoising steps (21 mm mean vertex error). Previous revision (ddpm_v2, the v2 recipe without temporal training): 594ce1af02d320dff3014418a309462345fa76b3.

v2-tf2-mismatch (2026-08-28) — this revision

The v2-tf2 recipe plus template-size mismatch augmentation (template_mismatch_range: [0.5, 2.0]): during training the rest template alone is rescaled by a log-uniform factor, so the template/observation size ratio stops being a constant 1 and the model must take scale from the observation, using the template for shape only. On the mismatched-template eval this revision is flat (~30 mm mean vertex error) across template scale factors 0.5–2 where the previous revision degrades 10–20x (160–250 mm at factor 1.88–2), and it ties the previous revision at true size (masked MSE 9.1e-4 vs 9.7e-4 at 5 denoising steps). On the real take-2 pants it tracks at metric scale — no hand-tuned template rescale — matching the previous revision's hand-tuned-scale arm (symmetric Chamfer 9.97 vs 9.61 mm). Training val 3.06e-4 m² sits above the previous revision's 2.38e-4 by design: the augmentation is train-only and val templates are always true-size. Previous revision (ddpm_v2_tf2, no mismatch augmentation): b0d762db67a12722817665a5365fa010bf589b4c.