dexgarmentlab-folding-lifting-dynamics-gps

GPSDynamicsModel — graph-based (GNN + Transformer) dynamics model for variable-vertex cloth meshes. Given the 3 previous mesh frames and a 3D gripper action, predicts the next 5 mesh frames via DDPM diffusion, using each cloth's own rest state and topology (no global template).

• Task data: DexGarmentLab mixed-garment fold + lift-place demos (dexgarmentlab_folding_lifting_meshes.h5, Cloth-splatters/dexgarmentlab-folding-lifting-meshes)
• Formulation: DDPM diffusion
• Max grippers: 1
• Cross-attention mode: parallel
• Max vertices per mesh: 2048
• Best validation loss: 0.00012608407087100204 (checkpoint in model/ is checkpoint-best)
• Training run: dexgarment_dyn_gps_2026-08-03_12-45-01_7138502 (full config in config.yml)

Cross-attention mode

cross_attn_mode: parallel. This checkpoint predates the option (added to the state-estimation model on 2026-05-13 and to the dynamics model on 2026-08-18) and was trained with the original parallel fusion of self- and cross-attention. The key was backfilled into model/config.json on 2026-08-19 so that from_pretrained cannot silently pick another mode — the modes share parameters, so a mismatch loads without error but runs a forward pass the model was never trained with.