Cloth-splatters/dexgarmentlab-lift-20260822-dynamics-gps-2grip

DexGarmentLab Lift 2026-08-22 — GPS two-gripper dynamics

GPSDynamicsModel trained on the full-state 2026-08-22 lift collection. Given the previous three mesh frames and the two hands' 3D actions, it predicts the next five mesh frames using DDPM diffusion.

• —Dataset: dexgarmentlab-lift-correspondence-20260822, file dexgarmentlab_lift_full_state_20260822.h5
• —Dataset SHA-256: 8eca09186491af5842ec7f09f28a3da885d25f6ba1eab6e74ead89126a977d40
• —Dataset: 1,800 episodes, 221 garments, 798,510 frames, garment-disjoint train/validation/test splits
• —Architecture: DDPM GPSDynamicsModel, 3 history frames, 5 predicted frames, max_grippers: 2
• —Training: seed 259, bf16, batch 32, cosine LR, history noise 1e-4
• —Best validation loss: 9.008239032937126e-06
• —Validation position MSE / zero-velocity baseline: 0.22321 (lower is better; the learned model is about 4.48x better)
• —Slurm job: 451440 (COMPLETED, 0:0)
• —Source branch: mesh-hypothesis-pf
• —Source revision at training/publication audit: f01f8091d052d77ef0b1e9edcea3e3b8683841b5
• —Weight SHA-256: d51598391abd1abf9f2d75a89299b7d007f7f0cbdbd68894be0c713c1d937770

Load through UniClothDiff's src.hub.resolve_checkpoint after checking out the source revision recorded in this card. The complete training recipe is config.yml; the Diffusers-compatible model is under model/.