Dataset card

README.md

@@ -0,0 +1,82 @@
+---
+license: mit
+task_categories:
+- robotics
+tags:
+- cloth
+- deformable-objects
+- garment
+- bimanual
+---
+
+# dexgarmentlab-dual-meshes-clean
+
+Two-hand DexGarmentLab garment manipulation demos with full mesh state,
+point-cloud observations, and per-hand gripper trajectories.
+
+- **File:** `dexgarmentlab_dual_gps2048_clean.h5` (1.4 GB)
+- **239 garments** — 136 Top, 77 Pants, 26 Dress (14 fine classes)
+- **894 trajectories**, **28 673 steps**, 12-59 steps per trajectory
+- **1637-2048 vertices**, <=12 102 bidirectional edges per mesh
+- **Tasks:** Fold (400 trajectories), Fling (249), Hang (200), Store (45)
+
+## Split
+
+The `training` / `validation` split is **baked into the file** at the top
+level and is **by garment**: 215 training cloths / 24 validation cloths, zero
+overlap. Every trajectory of a validation garment is held out, so this is a
+held-out-*object* protocol rather than the held-out-trajectory protocol the
+predecessor dataset forced. Loaders honour the groups directly and ignore
+`val_trajectory_ratio`.
+
+## Two hands
+
+644 of 894 trajectories move both hands; the rest move one and park the other.
+Per-hand truth:
+
+- `gripper_pos_2h` — `(2, 3)` per step, both end-effector positions
+- `actuated_vertices_2h` — `(2, V)` boolean, which vertices each hand holds
+- `hand_active` — trajectory attribute, `(2,)` boolean
+
+The file also carries legacy single-gripper `gripper_pos` `(3,)` /
+`actuated_vertices` `(V,)` for backwards compatibility. **On a two-handed
+trajectory these are lossy** and should not be used for training: the legacy
+point is a *virtual* gripper sitting between the hands, measured ~14 cm from
+the vertices it nominally drives (the real per-hand positions are 3-4 cm
+away), and the legacy mask covers only ~85 of the ~153 vertices the two hands
+actually hold.
+
+## `-clean`: training only
+
+Point clouds have had scene geometry (the robot arm) removed using the
+ground-truth cloth mesh — points further than `reject_m = 0.05` from the
+frame's own mesh are dropped. That segmentation uses information no estimator
+has at inference time, so this file is **correct for training and wrong for
+evaluating an observation model**. For evaluation use an uncleaned recording
+with an outlier-robust likelihood (the particle filter's
+`obs_trim_fraction`).
+
+## Structure
+
+```
+{training,validation}/<cloth_key>/
+    rest_positions      (V, 3) float32
+    edges               (E, 2) int32
+    faces               (F, 3) int32
+    trajectory_<i>/
+        actuated_vertices     (V,)    bool   # legacy, merged — see above
+        actuated_vertices_2h  (2, V)  bool
+        @hand_active          (2,)    bool
+        @task, @demo, @n_steps, @source
+        step_<nnnn>/
+            positions      (V, 3) float32
+            gripper_pos    (3,)   float32   # legacy, virtual — see above
+            gripper_pos_2h (2, 3) float32
+            pointclouds/   <= 2048 points, one camera
+```
+
+## Models trained on this dataset
+
+- [Cloth-splatters/dexgarmentlab-dual-dynamics-gps](https://huggingface.co/Cloth-splatters/dexgarmentlab-dual-dynamics-gps) — DDPM diffusion
+- [Cloth-splatters/dexgarmentlab-dual-dynamics-gps-flow](https://huggingface.co/Cloth-splatters/dexgarmentlab-dual-dynamics-gps-flow) — flow matching
+- [Cloth-splatters/dexgarmentlab-dual-dynamics-gns](https://huggingface.co/Cloth-splatters/dexgarmentlab-dual-dynamics-gns) — Graph Network Simulator