Upload folder using huggingface_hub

README.md

@@ -0,0 +1,67 @@
+---
+license: mit
+tags:
+  - robotics
+  - imitation-learning
+  - diffusion-policy
+  - cloth-manipulation
+---
+
+# DP3 floating-gripper garment policies — Hang_Trousers
+
+Inference-only checkpoints for five 3D Diffusion Policy variants trained on
+floating-gripper garment-hanging demonstrations in DexGarmentLab
+(`Hang_Trousers_FloatGrip`, 195 demos, 12,516 transitions).
+
+Code and full notes: **`floating-gripper-policy`** branch of
+<https://github.com/lei-shu/garmentlab>, under
+`DexGarmentLab/IL_Baselines/Diffusion_Policy_3D/` — see `EXPERIMENTS.md` for how
+these were trained and `SIM_EVAL_HANDOFF.md` for how to run them.
+
+## The five variants
+
+Each differs only in the observation representation and the point encoder.
+`EE MAE` is held-out end-effector position error against demo actions — **not**
+task success, which has not been measured yet.
+
+| file | encoder | observation | edges | epoch | EE MAE |
+|---|---|---|---|---|---|
+| `trousers_pn_depth.ckpt` | PointNet | depth cloud 2048×3 | — | 550 | **6.25 mm** |
+| `trousers_gnn_depth.ckpt` | mesh-GNN | depth cloud | kNN | 200 | 6.85 mm |
+| `trousers_pn_mesh.ckpt` | PointNet | decimated mesh + pothook | — | 900 | 6.94 mm |
+| `trousers_gnn_mesh.ckpt` | mesh-GNN | mesh nodes | **true garment faces** | 300 | 7.24 mm |
+| `trousers_gnn_knn.ckpt` | mesh-GNN | mesh nodes | kNN | 200 | 7.35 mm |
+
+All five sit within ~1 mm of each other on a 4-episode validation split, so treat
+them as near-indistinguishable offline. They do beat an observation-free
+"hold position" baseline (48.9 mm) by 8×, so they are genuinely using the
+observation.
+
+## Loading
+
+```python
+import torch, dill, hydra
+from omegaconf import OmegaConf
+OmegaConf.register_new_resolver("eval", eval, replace=True)
+
+p      = torch.load("trousers_pn_depth.ckpt", map_location="cpu", pickle_module=dill)
+policy = hydra.utils.instantiate(p["cfg"].policy)
+policy.load_state_dict(p["state_dicts"]["ema_model"])
+policy.cuda().eval()
+
+# obs: point_cloud (B, 3, 2048, 3) + agent_pos (B, 3, 8)
+# out: (B, 4, 8) = [left_xyz, right_xyz, grip_L, grip_R], absolute metres
+action = policy.predict_action(obs)["action"]
+```
+
+These carry EMA weights only (~0.98 GiB each); the training checkpoints they came
+from were 3.91 GiB with optimizer state. The normalizer travels inside the
+`state_dict`, so actions come back in metres — do not rescale. Gripper bits are
+regressed continuously; threshold at 0.5.
+
+`trousers_gnn_mesh` additionally needs `edges`, `n_nodes`, `n_edges` in the
+observation dict (see `SIM_EVAL_HANDOFF.md` §4a — the stock `RobotRunner` drops
+them and it will raise `KeyError: 'edges'`).
+
+Regenerate or re-upload with
+`python tools/export_checkpoints.py --out export/ --push <repo>`.

manifest.json

@@ -0,0 +1,27 @@
+{
+ "trousers_pn_depth": {
+  "file": "trousers_pn_depth.ckpt",
+  "epoch": 550,
+  "gib": 0.98
+ },
+ "trousers_pn_mesh": {
+  "file": "trousers_pn_mesh.ckpt",
+  "epoch": 900,
+  "gib": 0.98
+ },
+ "trousers_gnn_depth": {
+  "file": "trousers_gnn_depth.ckpt",
+  "epoch": 200,
+  "gib": 0.98
+ },
+ "trousers_gnn_knn": {
+  "file": "trousers_gnn_knn.ckpt",
+  "epoch": 200,
+  "gib": 0.98
+ },
+ "trousers_gnn_mesh": {
+  "file": "trousers_gnn_mesh.ckpt",
+  "epoch": 300,
+  "gib": 0.98
+ }
+}

trousers_gnn_depth.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d960c79cd3f498a51abbfdd1d7a93bf05d66804269979499b6afbdae5badc45
+size 1050205214

trousers_gnn_knn.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db32cd299db3ad583f98e68e477005eccd71e1efbd180b4b4d1910a7b8843561
+size 1050204694

trousers_gnn_mesh.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8228e86ea8937ac0299b9b1333c7e42d519af3cd2fae1ca3123a6ff7e84dd27d
+size 1050204954

trousers_pn_depth.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bacf2432b71f6914c4f2636814f31d0b09fec3e76dd681b5db09697700ebbc7
+size 1049741498

trousers_pn_mesh.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a939d28ab6bca9ca820d092e1601002cae9d5f4fbae81004a8750eca031ab64
+size 1049741306