Describe native HDF5 release and disclose repeated trajectories

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+manifests/episodes.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,4 +1,124 @@
 ---
 license: cc-by-nc-4.0
----nse: apache-2.0
+language:
+- en
+task_categories:
+- robotics
+tags:
+- libero
+- libero-pro2
+- robot-learning
+- imitation-learning
+- vision-language-action
+- hdf5
+pretty_name: LIBERO-Pro2.0 Native HDF5 Training Data
+size_categories:
+- 10K<n<100K
 ---
+
+# LIBERO-Pro2.0 — Native HDF5 Training Data
+
+**Release status: UPLOAD IN PROGRESS.** Check `release.json`; the complete release is available only when its status is `complete`.
+
+This repository distributes the native HDF5 training demonstrations for LIBERO-Pro2.0, with per-instance metadata, trajectory source indexes, and review videos. This is the HDF5 release, not an RLDS or LeRobot conversion. No simulator assets, model weights, or evaluation rollouts are bundled here.
+
+## Dataset scope
+
+| Suite | Base task IDs included | Instances | Stored demonstrations |
+|---|---|---:|---:|
+| libero_spatial | 1, 2, 3, 4, 5, 6, 7, 9 | 338 | 10,140 |
+| libero_object | 0, 2, 3, 4, 5, 6, 7, 9 | 344 | 10,320 |
+| libero_goal | 0, 1, 2, 4, 5, 7, 8, 9 | 323 | 9,690 |
+| libero_10 | 0, 1, 2, 3, 4, 6, 7, 9 | 344 | 10,320 |
+| Total | 32 base tasks | 1,349 | 40,470 |
+
+The source audit reports **8,258,504 total steps**. Each instance contains 30 stored demonstrations and 5 review videos (6,745 videos total). BASE and perturbation instances are included; case coverage is not identical for every task. Exact statistics and per-episode records are in `manifests/statistics.json` and `manifests/episodes.jsonl`.
+
+The source package was collected in September 2026. The original archive, approximately 218.8 GiB, passed both per-part and concatenated-archive SHA-256 verification before release preparation. Production HDF5 files, metadata, source indexes, and review videos are preserved byte-for-byte. Pilot HDF5 files, collection logs, progress files, attempt logs, and lock files are excluded.
+
+## Trajectory provenance and repeated demonstrations
+
+**40,470 stored demonstrations are not 40,470 independent source trajectories.** The supplied source audit reports 37,396 unique source trajectories: 1,243 instances have 30 distinct sources, while **106 instances replicate one independently validated successful source trajectory into 30 stored demonstrations**. These copies are retained to preserve the original data; they must not be interpreted as independent samples.
+
+`manifests/fanout_instances.json` lists these instances. `manifests/episodes.jsonl` provides `keep_without_declared_fanout`: selecting true keeps all demonstrations from ordinary instances and one representative from each declared fanout instance (37,396 records). This flag removes declared fanout copies only. Release preparation does not perform a new global content-based trajectory deduplication. Some original source indexes do not contain fingerprints; missing values are explicitly null.
+
+The data were generated using a privileged-state oracle, rather than human teleoperation. The supplied metadata marks oracle privileged state as not recorded in observations. HDF5 attributes and source indexes nevertheless contain provenance and result information; training code should explicitly select observation/action fields. Review videos are successful-trajectory examples, not unbiased evaluation results. Do not use random episode splits across replicated trajectories to claim independent held-out performance. This release supplies training data only, not a new validation or test split.
+
+## Layout
+
+```text
+data/<suite>/task_<id>/<case>/
+  episodes.hdf5
+  metadata.json
+  source_index.json
+  review_videos/success_000.mp4  # through success_004.mp4
+manifests/
+  statistics.json
+  episodes.jsonl
+  fanout_instances.json
+  files.jsonl
+  SHA256SUMS
+  source_final_audit.json
+  archive_checksums.json
+  schema_example.json
+  release_validation.json
+scripts/inspect_episode.py
+release.json
+```
+
+Paths inside original `source_index.json` are relative to this repository's `data/` directory; the suffix after `::` identifies the HDF5 group. `review_videos` references are relative to the instance directory. BDDL paths and manifest hashes identify the original benchmark configuration; the referenced simulator/BDDL assets are not included in this data-only release.
+
+## Native schema
+
+Each file contains `data/demo_0` through `data/demo_29`. `T` varies by demonstration. The inspected example exposes the following fields; inspect original attributes and metadata before integrating a controller.
+
+| Field under a demo | Shape | Type / meaning |
+|---|---|---|
+| actions | `(T, 7)` | float32, relative control commands |
+| obs/agentview_image | `(T, 128, 128, 3)` | uint8 RGB, scene camera |
+| obs/robot0_eye_in_hand_image | `(T, 128, 128, 3)` | uint8 RGB, wrist camera |
+| obs/robot0_eef_pos | `(T, 3)` | float32 end-effector position |
+| obs/robot0_eef_quat_xyzw | `(T, 4)` | float32 quaternion, xyzw order |
+| obs/robot0_gripper_qpos | `(T, 2)` | float32 gripper joint positions |
+| obs/robot0_joint_pos | `(T, 7)` | float32 robot joint positions |
+| obs/robot0_joint_vel | `(T, 7)` | float32 robot joint velocities |
+| rewards | `(T,)` | float32 |
+| dones | `(T,)` | bool |
+| server_step_seconds | `(T,)` | recorded server timing, where present |
+
+Dataset attributes include `env_args`, `generation_metadata`, and `total`; demo attributes include success and source metadata. See `manifests/schema_example.json` for the original action schema from an inspected episode, including control frequency, controller, and gripper semantics. Images and actions have not been rotated, resized, normalized, or converted for publication. This repository does not supply training normalization statistics or claim drop-in compatibility with arbitrary policy loaders.
+
+## Download and inspect
+
+```bash
+pip install -U huggingface_hub h5py
+
+# Download one suite first. Set the official endpoint if your environment uses a mirror.
+HF_ENDPOINT=https://huggingface.co hf download zhouxueyang/LIBERO-Pro2.0 \
+  --repo-type dataset --local-dir ./LIBERO-Pro2.0 \
+  --include 'data/libero_spatial/**' --include 'manifests/**' \
+  --include 'scripts/**' --include 'README.md' --include 'release.json'
+
+python ./LIBERO-Pro2.0/scripts/inspect_episode.py \
+  ./LIBERO-Pro2.0/data/libero_spatial/task_1/BASE/episodes.hdf5
+
+# Full dataset download (requires approximately 219 GiB plus cache/headroom).
+HF_ENDPOINT=https://huggingface.co hf download zhouxueyang/LIBERO-Pro2.0 \
+  --repo-type dataset --local-dir ./LIBERO-Pro2.0
+
+# For a complete local download; subset downloads omit some listed files.
+cd LIBERO-Pro2.0
+sha256sum -c manifests/SHA256SUMS
+```
+
+Raw HDF5 is intended for local training loaders. This release does not advertise automatic Hugging Face Dataset Viewer, RLDS, or LeRobot compatibility. Those would require separately validated conversions.
+
+## Validation and limitations
+
+The source audit is preserved as `manifests/source_final_audit.json`. Release preparation additionally checks every production HDF5 file for readability, demo counts, action/image shapes, reward/done lengths, success attributes, and step totals; it hashes every distributed payload file. It does not replay physics, decode every image/video, or independently re-establish the source audit's uniqueness and task-success claims. A completed release is marked only after its remote file sizes and content hashes match the prepared payload.
+
+## License
+
+This dataset repository uses **CC BY-NC 4.0**, retaining the license selected by the repository owner. See https://creativecommons.org/licenses/by-nc/4.0/ . This metadata does not relicense external simulator code, benchmark assets, or other third-party materials. Refer to their original licenses when using them separately.
+
+No paper DOI or formal citation is asserted for this release. Cite this repository URL and the specific revision used; benchmark-code and publication references can be added by the maintainers.

manifests/archive_checksums.json

@@ -0,0 +1,38 @@
+{
+  "archive_sha256": "94d9b07bf1544818579be86fd8e928c958b6cd8bbccb97a68fe4cd377fcbf7e6",
+  "parts": [
+    {
+      "sha256": "1ce75c7785218f14c13a64611c054016a07d35568f36f6ce85e5e694ce7ca931",
+      "filename": "training_data.tar.part-001"
+    },
+    {
+      "sha256": "c4966cecb75c4cdae9201a4586e986d56f60990dd1656011c43028d46ccea0d8",
+      "filename": "training_data.tar.part-002"
+    },
+    {
+      "sha256": "5e7c07607447a3ced40d6878b5ef26512e1b7e13fc4a4cc80124d69cf3cc4072",
+      "filename": "training_data.tar.part-003"
+    },
+    {
+      "sha256": "bddf2094172283c39870d516a56f4aa6b2ecec8e82dc1356c2325324d2e3f9d5",
+      "filename": "training_data.tar.part-004"
+    },
+    {
+      "sha256": "64c53b374e76fc40b8ea36a86e5a8f1a5f72fcb1d141a7b5a2825c43a9e0db8b",
+      "filename": "training_data.tar.part-005"
+    },
+    {
+      "sha256": "2801e5ec9a4c0923c0f44c47d78b16b98d8eb5405f2254dc7eb01e3d04654f62",
+      "filename": "training_data.tar.part-006"
+    },
+    {
+      "sha256": "f1a0cf2bc5572438c3adf6353d21919aca980cc45fd0d438a4053aaab69c9ae3",
+      "filename": "training_data.tar.part-007"
+    },
+    {
+      "sha256": "db5b83c77484737c75a21ba39dff6dd95274eca8d13dd7901c1aedd663add5a6",
+      "filename": "training_data.tar.part-008"
+    }
+  ],
+  "verified_before_extraction": true
+}

manifests/episodes.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:373a3203923638884c1a20814a7ed9e0e509e2441ffea317df894f27a11779bb
+size 14821608

manifests/fanout_instances.json

@@ -0,0 +1,108 @@
+[
+  "libero_object/task_2/S06",
+  "libero_object/task_2/D03",
+  "libero_object/task_3/D03",
+  "libero_object/task_4/S06",
+  "libero_object/task_4/D03",
+  "libero_object/task_5/D03",
+  "libero_object/task_7/D03",
+  "libero_object/task_9/S06",
+  "libero_object/task_9/D03",
+  "libero_object/task_6/S06",
+  "libero_object/task_6/D01",
+  "libero_object/task_6/D03",
+  "libero_object/task_6/D05",
+  "libero_spatial/task_1/D03",
+  "libero_spatial/task_1/D18",
+  "libero_spatial/task_2/D03",
+  "libero_spatial/task_3/S05",
+  "libero_spatial/task_3/D03",
+  "libero_spatial/task_3/D05",
+  "libero_spatial/task_4/S05",
+  "libero_spatial/task_4/D03",
+  "libero_spatial/task_4/D05",
+  "libero_spatial/task_4/D15",
+  "libero_spatial/task_4/D18",
+  "libero_spatial/task_5/D18",
+  "libero_spatial/task_6/D01",
+  "libero_spatial/task_6/D05",
+  "libero_spatial/task_6/D17",
+  "libero_spatial/task_6/D18",
+  "libero_spatial/task_7/D18",
+  "libero_spatial/task_9/S05",
+  "libero_spatial/task_9/D01",
+  "libero_spatial/task_9/D05",
+  "libero_spatial/task_9/D17",
+  "libero_spatial/task_9/D18",
+  "libero_goal/task_0/D01",
+  "libero_goal/task_5/S01",
+  "libero_goal/task_5/S05",
+  "libero_goal/task_5/D01",
+  "libero_goal/task_7/D01",
+  "libero_goal/task_9/S06",
+  "libero_10/task_0/S09",
+  "libero_10/task_0/D03",
+  "libero_10/task_0/D05",
+  "libero_10/task_0/D18",
+  "libero_10/task_1/S06",
+  "libero_10/task_1/D05",
+  "libero_10/task_1/D18",
+  "libero_10/task_2/S06",
+  "libero_10/task_3/S01",
+  "libero_10/task_3/S02",
+  "libero_10/task_3/S03",
+  "libero_10/task_3/S04",
+  "libero_10/task_3/S05",
+  "libero_10/task_3/S06",
+  "libero_10/task_3/S07",
+  "libero_10/task_3/S09",
+  "libero_10/task_3/S10",
+  "libero_10/task_3/S12",
+  "libero_10/task_3/D01",
+  "libero_10/task_3/D03",
+  "libero_10/task_3/D04",
+  "libero_10/task_3/D05",
+  "libero_10/task_3/D06",
+  "libero_10/task_3/D07",
+  "libero_10/task_3/D08",
+  "libero_10/task_3/D09",
+  "libero_10/task_3/D11",
+  "libero_10/task_3/D12",
+  "libero_10/task_3/D13",
+  "libero_10/task_3/D14",
+  "libero_10/task_3/D15",
+  "libero_10/task_3/D16",
+  "libero_10/task_3/D17",
+  "libero_10/task_3/D18",
+  "libero_10/task_4/S05",
+  "libero_10/task_4/S20",
+  "libero_10/task_4/D05",
+  "libero_10/task_6/S05",
+  "libero_10/task_6/D03",
+  "libero_10/task_6/D05",
+  "libero_10/task_7/S01",
+  "libero_10/task_7/S02",
+  "libero_10/task_7/S03",
+  "libero_10/task_7/S04",
+  "libero_10/task_7/D01",
+  "libero_10/task_7/D05",
+  "libero_10/task_7/D14",
+  "libero_10/task_7/D15",
+  "libero_10/task_7/D16",
+  "libero_10/task_7/D17",
+  "libero_10/task_9/S01",
+  "libero_10/task_9/S02",
+  "libero_10/task_9/S03",
+  "libero_10/task_9/S04",
+  "libero_10/task_9/S05",
+  "libero_10/task_9/S06",
+  "libero_10/task_9/S07",
+  "libero_10/task_9/S09",
+  "libero_10/task_9/D01",
+  "libero_10/task_9/D03",
+  "libero_10/task_9/D05",
+  "libero_10/task_9/D13",
+  "libero_10/task_9/D14",
+  "libero_10/task_9/D17",
+  "libero_10/task_9/D18"
+]

manifests/source_final_audit.json

@@ -0,0 +1,46 @@
+{
+  "artifacts": {
+    "metadata_complete": 1349,
+    "residual_oracle_shard_directories": 0,
+    "review_videos": 6745,
+    "review_videos_per_instance": 5,
+    "source_indexes_validated": 1349
+  },
+  "audit_duration_seconds": 60.154,
+  "audited_at": "2026-09-21T07:34:23.159435+00:00",
+  "complete": true,
+  "errors": [],
+  "hdf5": {
+    "action_shape_required": "(steps, 7)",
+    "dataset_bytes": 233487980473,
+    "datasets_validated": 1349,
+    "demos_validated": 40470,
+    "image_shape_required": "(steps, 128, 128, 3)",
+    "terminal_success_required": true,
+    "total_steps": 8258504
+  },
+  "manifest_sha256": "71a6e4e25a2a4a162844d77136e7bff8e6f69c839d2d8194620c4afeabde2bba",
+  "manifest_version": "libero-pro2.0-1706-v1",
+  "progress": {
+    "current_instance": null,
+    "instances_completed": 1349,
+    "instances_total": 1349,
+    "trajectories_completed": 40470,
+    "trajectories_per_instance": 30,
+    "trajectories_target": 40470,
+    "updated_at": "2026-09-21T07:20:03.237804+00:00"
+  },
+  "requirements": {
+    "instances_target": 1349,
+    "review_videos_per_instance": 5,
+    "trajectories_per_instance": 30,
+    "trajectories_target": 40470
+  },
+  "schema_version": "libero_pro_final_audit_v1",
+  "uniqueness": {
+    "fanout_policy": "one independently validated successful source trajectory copied to 30 demos",
+    "single_success_fanout_instances": 106,
+    "strict_unique_instances": 1243,
+    "unique_source_trajectories": 37396
+  }
+}

manifests/statistics.json

@@ -0,0 +1,76 @@
+{
+  "suites": {
+    "libero_spatial": {
+      "tasks": [
+        "task_1",
+        "task_2",
+        "task_3",
+        "task_4",
+        "task_5",
+        "task_6",
+        "task_7",
+        "task_9"
+      ],
+      "instances": 338,
+      "demonstrations": 10140,
+      "steps": 1722293,
+      "bytes": 46043858051
+    },
+    "libero_object": {
+      "tasks": [
+        "task_0",
+        "task_2",
+        "task_3",
+        "task_4",
+        "task_5",
+        "task_6",
+        "task_7",
+        "task_9"
+      ],
+      "instances": 344,
+      "demonstrations": 10320,
+      "steps": 2003300,
+      "bytes": 65008820801
+    },
+    "libero_goal": {
+      "tasks": [
+        "task_0",
+        "task_1",
+        "task_2",
+        "task_4",
+        "task_5",
+        "task_7",
+        "task_8",
+        "task_9"
+      ],
+      "instances": 323,
+      "demonstrations": 9690,
+      "steps": 1381059,
+      "bytes": 35242360519
+    },
+    "libero_10": {
+      "tasks": [
+        "task_0",
+        "task_1",
+        "task_2",
+        "task_3",
+        "task_4",
+        "task_6",
+        "task_7",
+        "task_9"
+      ],
+      "instances": 344,
+      "demonstrations": 10320,
+      "steps": 3151852,
+      "bytes": 88398479966
+    }
+  },
+  "instances": 1349,
+  "demonstrations": 40470,
+  "source_audit_unique_trajectories": 37396,
+  "declared_fanout_instances": 106,
+  "steps": 8258504,
+  "payload_files": 10792,
+  "payload_bytes": 234693519337,
+  "uniqueness_note": "Source audit reports 37396 unique source trajectories. The keep flag removes declared fanout copies only; no new global trajectory-content deduplication was performed."
+}

release.json

@@ -0,0 +1,9 @@
+{
+  "status": "uploading",
+  "format": "native-hdf5",
+  "version": "1.0.0",
+  "expected_instances": 1349,
+  "expected_payload_files": 10792,
+  "expected_payload_bytes": 234693519337,
+  "note": "Upload in progress. Wait for status complete before using the full release."
+}

scripts/inspect_episode.py

@@ -0,0 +1,17 @@
+"""Read a native LIBERO-Pro2.0 demonstration without executing simulator code."""
+import argparse
+import json
+import h5py
+
+p=argparse.ArgumentParser()
+p.add_argument('hdf5')
+p.add_argument('--demo',default='demo_0')
+a=p.parse_args()
+with h5py.File(a.hdf5,'r') as f:
+    d=f['data'][a.demo]
+    print('actions:',d['actions'].shape,d['actions'].dtype)
+    for key in d['obs']:
+        print('obs/'+key,d['obs'][key].shape,d['obs'][key].dtype)
+    print('success:',bool(d.attrs.get('success',False)))
+    if 'source_metadata' in d.attrs:
+        print(json.dumps(json.loads(d.attrs['source_metadata']),indent=2))