Datasets:

zhouxueyang
/

LIBERO-Pro2.0-Train

@@ -1,132 +1,146 @@
----
-license: cc-by-nc-4.0
-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
-configs:
-- config_name: episodes
-  default: true
-  data_files:
-  - split: train
-    path: viewer/episodes/*/*.parquet
-- config_name: libero_10
-  data_files:
-  - split: train
-    path: viewer/episodes/libero_10/*.parquet
-- config_name: libero_goal
-  data_files:
-  - split: train
-    path: viewer/episodes/libero_goal/*.parquet
-- config_name: libero_object
-  data_files:
-  - split: train
-    path: viewer/episodes/libero_object/*.parquet
-- config_name: libero_spatial
-  data_files:
-  - split: train
-    path: viewer/episodes/libero_spatial/*.parquet
-- config_name: review_videos
-  data_files:
-  - split: train
-    path: viewer/review_videos/*.parquet
----
-<h1 align="center">
-LIBERO-Pro2.0: A Benchmark for Evaluating Robust Embodied Intelligence
-</h1>
-<p align="center">
-  🌐 <a href="https://racingemperor.github.io/LIBERO-Pro2.0/index.html"><strong>Website</strong></a> |
-  📖 <a href="https://racingemperor.github.io/LIBERO-Pro2.0/docs.html"><strong>Docs</strong></a> |
-  🏆 <a href="https://racingemperor.github.io/LIBERO-Pro2.0/leaderboard.html"><strong>Leaderboard</strong></a> |
-  🤗 <a href="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/viewer/episodes/train"><strong>Dataset</strong></a> |
-  🎬 <a href="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/viewer/review_videos/train"><strong>Videos</strong></a>
-</p>
-<p>
-  <img src="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/resolve/main/assets/card/static-pose.webp" alt="Static shift: base and perturbed target placement" width="100%">
-  <img src="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/resolve/main/assets/card/dynamic-occlusion.webp" alt="Runtime intervention: target visible, temporarily occluded, then visible again" width="100%">
-</p>
-*Examples from LIBERO-Pro2.0: target pose variation at reset (top) and temporary occlusion during execution (bottom).*
-## 🔥 Overview
-This repository hosts the **native HDF5 training data for LIBERO-Pro2.0**, a benchmark for studying robust robot manipulation under changing conditions. The benchmark introduces **22 static shifts** and **20 dynamic interventions** to test generalization and recovery while preserving task goals.
-The training data contains **40,470 stored demonstrations** across **32 base tasks** in four LIBERO suites. Each demonstration includes a language instruction, scene and wrist camera observations, robot states, and actions.
-## 🚀 Dataset Highlights
-- **Static and dynamic variation:** demonstrations cover base tasks and perturbation variants across 1,349 task/variation instances.
-- **Multimodal trajectories:** 8,258,504 steps with two 128 × 128 RGB camera views, robot states, and 7-dimensional actions.
-- **Browsable demonstrations:** one Viewer row per episode, plus 6,745 review videos showing successful examples.
-## 📊 LIBERO-Pro2.0 Benchmark
-The benchmark organizes perturbations into six domains:
-1. **Condition** — Task geometry, object poses, and interaction constraints.
-2. **Environment** — Scene composition, appearance, and visual context.
-3. **Observation** — Image formation, sensor quality, and observation timing.
-4. **Execution** — Control timing, command delivery, and action execution.
-5. **Robot** — Robot state, calibration, and recoverable execution outcomes.
-6. **Language** — Equivalent instructions that preserve the original task goal.
-See the [benchmark documentation](https://racingemperor.github.io/LIBERO-Pro2.0/docs.html#taxonomy) for perturbation designs and the [leaderboard](https://racingemperor.github.io/LIBERO-Pro2.0/leaderboard.html) for evaluation results.
-## 📁 Training Data
-| Suite | Base tasks | Instances | Stored demonstrations |
-|---|---:|---:|---:|
-| LIBERO-Spatial | 8 | 338 | 10,140 |
-| LIBERO-Object | 8 | 344 | 10,320 |
-| LIBERO-Goal | 8 | 323 | 9,690 |
-| LIBERO-10 | 8 | 344 | 10,320 |
-| **Total** | **32** | **1,349** | **40,470** |
-Trajectories are oracle-generated. This repository provides training data only, with no validation or test split. Full observation and action sequences are stored in HDF5; the Viewer provides camera previews and the corresponding file/group for each episode.
-**Trajectory provenance:** the source audit reports **37,396 unique source trajectories**. In 106 instances, one source trajectory is replicated into 30 stored demonstrations. Select `keep_without_declared_fanout = true` in the episode index to exclude declared copies. See the [data format guide](docs/data_format.md#trajectory-provenance) for details.
-## 🛠️ Download
-```bash
-pip install -U huggingface_hub
-hf download zhouxueyang/LIBERO-Pro2.0 \
-  --repo-type dataset --local-dir ./LIBERO-Pro2.0
-```
-Allow approximately **220 GiB**, plus cache space. To download one suite, add `--include "data/libero_spatial/**"`.
-## 🔧 Usage
-Install `h5py` and read a trajectory:
-```python
-import h5py
-path = "LIBERO-Pro2.0/data/libero_spatial/task_1/BASE/episodes.hdf5"
-with h5py.File(path, "r") as f:
-    episode = f["data/demo_0"]
-    actions = episode["actions"][:]
-    scene_images = episode["obs/agentview_image"][:]
-    wrist_images = episode["obs/robot0_eye_in_hand_image"][:]
-```
-Refer to the [data format guide](docs/data_format.md) for the directory structure, observation fields, action metadata, and episode index.
-## License
-This dataset is licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Please cite this repository and the revision used. External simulator code and benchmark assets retain their respective licenses.

+---
+license: cc-by-nc-4.0
+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
+configs:
+- config_name: episodes
+  default: true
+  data_files:
+  - split: train
+    path: viewer/episodes/*/*.parquet
+- config_name: libero_10
+  data_files:
+  - split: train
+    path: viewer/episodes/libero_10/*.parquet
+- config_name: libero_goal
+  data_files:
+  - split: train
+    path: viewer/episodes/libero_goal/*.parquet
+- config_name: libero_object
+  data_files:
+  - split: train
+    path: viewer/episodes/libero_object/*.parquet
+- config_name: libero_spatial
+  data_files:
+  - split: train
+    path: viewer/episodes/libero_spatial/*.parquet
+- config_name: review_videos
+  data_files:
+  - split: train
+    path: viewer/review_videos/*.parquet
+---
+<h1 align="center">
+LIBERO-Pro2.0: A Benchmark for Evaluating Robust Embodied Intelligence
+</h1>
+<p align="center">
+  🌐 <a href="https://racingemperor.github.io/LIBERO-Pro2.0/index.html"><strong>Website</strong></a> |
+  📖 <a href="https://racingemperor.github.io/LIBERO-Pro2.0/docs.html"><strong>Docs</strong></a> |
+  🏆 <a href="https://racingemperor.github.io/LIBERO-Pro2.0/leaderboard.html"><strong>Leaderboard</strong></a> |
+  🤗 <a href="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/viewer/episodes/train"><strong>Dataset</strong></a> |
+  🎬 <a href="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/viewer/review_videos/train"><strong>Videos</strong></a>
+</p>
+<p>
+  <img src="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/resolve/main/assets/card/static-pose.webp" alt="Static shift: base and perturbed target placement" width="100%">
+  <img src="https://huggingface.co/datasets/zhouxueyang/LIBERO-Pro2.0/resolve/main/assets/card/dynamic-occlusion.webp" alt="Runtime intervention: target visible, temporarily occluded, then visible again" width="100%">
+</p>
+*Examples from LIBERO-Pro2.0: target pose variation at reset (top) and temporary occlusion during execution (bottom).*
+## 🔥 Overview
+This repository hosts the **native HDF5 training data for LIBERO-Pro2.0**, a benchmark for studying robust robot manipulation under changing conditions. The benchmark introduces **22 static shifts** and **20 dynamic interventions** to test generalization and recovery while preserving task goals.
+The training data contains **40,470 stored demonstrations** across **32 base tasks** in four LIBERO suites. Each demonstration includes a language instruction, scene and wrist camera observations, robot states, and actions.
+## 🚀 Dataset Highlights
+- **Static and dynamic variation:** demonstrations cover base tasks and perturbation variants across 1,349 task/variation instances.
+- **Multimodal trajectories:** 8,258,504 steps with two 128 × 128 RGB camera views, robot states, and 7-dimensional actions.
+- **Browsable demonstrations:** one Viewer row per episode, plus 6,745 review videos showing successful examples.
+## 📊 LIBERO-Pro2.0 Benchmark
+The benchmark organizes perturbations into six domains:
+1. **Condition** — Task geometry, object poses, and interaction constraints.
+2. **Environment** — Scene composition, appearance, and visual context.
+3. **Observation** — Image formation, sensor quality, and observation timing.
+4. **Execution** — Control timing, command delivery, and action execution.
+5. **Robot** — Robot state, calibration, and recoverable execution outcomes.
+6. **Language** — Equivalent instructions that preserve the original task goal.
+See the [benchmark documentation](https://racingemperor.github.io/LIBERO-Pro2.0/docs.html#taxonomy) for perturbation designs and the [leaderboard](https://racingemperor.github.io/LIBERO-Pro2.0/leaderboard.html) for evaluation results.
+## 📁 Training Data
+| Suite | Base tasks | Instances | Stored demonstrations |
+|---|---:|---:|---:|
+| LIBERO-Spatial | 8 | 338 | 10,140 |
+| LIBERO-Object | 8 | 344 | 10,320 |
+| LIBERO-Goal | 8 | 323 | 9,690 |
+| LIBERO-10 | 8 | 344 | 10,320 |
+| **Total** | **32** | **1,349** | **40,470** |
+Trajectories are oracle-generated. This repository provides training data only, with no validation or test split. Full observation and action sequences are stored in HDF5; the Viewer provides camera previews and the corresponding file/group for each episode.
+**Trajectory provenance:** the source audit reports **37,396 unique source trajectories**. In 106 instances, one source trajectory is replicated into 30 stored demonstrations. Select `keep_without_declared_fanout = true` in the episode index to exclude declared copies. See the [data format guide](docs/data_format.md#trajectory-provenance) for details.
+## Simulation Assets
+The simulation assets used by the **1,688-task LIBERO-Pro release** are also
+available as [assets.zip](assets.zip). The archive contains objects, scenes,
+textures, and perturbation assets; it is separate from the HDF5 training data.
+Download only this archive and its verification files if you are setting up
+evaluation. See the [asset installation guide](docs/simulation_assets.md).
+BDDL task definitions and evaluation code are maintained in the
+[GitHub main branch](https://github.com/Zxy-MLlab/LIBERO-Pro2.0/tree/main).
+The asset bundle is pinned to source commit
+`3ac451e6e85ee6995bf684b56f6c04f8ef7c8829`; it does not change the training split,
+stored trajectories, or Dataset Viewer.
+## 🛠️ Download
+```bash
+pip install -U huggingface_hub
+hf download zhouxueyang/LIBERO-Pro2.0 \
+  --repo-type dataset --local-dir ./LIBERO-Pro2.0
+```
+Allow approximately **220 GiB**, plus cache space. To download one suite, add `--include "data/libero_spatial/**"`.
+## 🔧 Usage
+Install `h5py` and read a trajectory:
+```python
+import h5py
+path = "LIBERO-Pro2.0/data/libero_spatial/task_1/BASE/episodes.hdf5"
+with h5py.File(path, "r") as f:
+    episode = f["data/demo_0"]
+    actions = episode["actions"][:]
+    scene_images = episode["obs/agentview_image"][:]
+    wrist_images = episode["obs/robot0_eye_in_hand_image"][:]
+```
+Refer to the [data format guide](docs/data_format.md) for the directory structure, observation fields, action metadata, and episode index.
+## License
+This dataset is licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Please cite this repository and the revision used. External simulator code and benchmark assets retain their respective licenses.

assets.zip ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:c2a160e5b8ad4552a345adb54551f33bbce74c0cf9f79218a588197ba8eaa9d0
+size 483197257

assets.zip.sha256 ADDED Viewed

	@@ -0,0 +1 @@


1	+ c2a160e5b8ad4552a345adb54551f33bbce74c0cf9f79218a588197ba8eaa9d0 assets.zip

docs/simulation_assets.md ADDED Viewed

	@@ -0,0 +1,92 @@

+# LIBERO-Pro simulation assets
+This archive supplies the environment assets for the 1,688-task release of
+LIBERO-Pro. It contains the existing LIBERO objects, scenes, textures, and the
+additional perturbation assets. BDDL files, initial states, and evaluation code
+remain in the [GitHub main branch](https://github.com/Zxy-MLlab/LIBERO-Pro2.0/tree/main).
+The HDF5 training data in `data/` and the Dataset Viewer are unchanged.
+## Download
+The ZIP is approximately 483 MB and expands to approximately 821 MB. Download
+only the asset supplement:
+```bash
+pip install -U huggingface_hub
+hf download zhouxueyang/LIBERO-Pro2.0 --repo-type dataset \
+  --include "assets.zip" "assets.zip.sha256" "manifests/assets_manifest.json" \
+            "scripts/verify_assets.py" "docs/simulation_assets.md" \
+  --local-dir ./libero-pro-assets
+python ./libero-pro-assets/scripts/verify_assets.py \
+  --archive ./libero-pro-assets/assets.zip \
+  --manifest ./libero-pro-assets/manifests/assets_manifest.json
+```
+## Install into the current evaluation release
+The compatible source commit is `3ac451e6e85ee6995bf684b56f6c04f8ef7c8829`.
+Use a fresh checkout so existing environments and running jobs are unaffected.
+The GitHub repository currently requires access from its owner.
+```bash
+GIT_LFS_SKIP_SMUDGE=1 git clone --branch main \
+  https://github.com/Zxy-MLlab/LIBERO-Pro2.0.git ./LIBERO-Pro-eval
+git -C ./LIBERO-Pro-eval checkout 3ac451e6e85ee6995bf684b56f6c04f8ef7c8829
+python -m zipfile -e ./libero-pro-assets/assets.zip \
+  ./LIBERO-Pro-eval/simulator-source/libero/libero
+python ./libero-pro-assets/scripts/verify_assets.py \
+  --assets-dir ./LIBERO-Pro-eval/simulator-source/libero/libero/assets \
+  --manifest ./libero-pro-assets/manifests/assets_manifest.json
+```
+The archive already contains the `assets/` directory. Extract into its parent,
+not into another `assets/` directory:
+```text
+LIBERO-Pro-eval/
+└── simulator-source/libero/libero/
+    └── assets/
+        ├── articulated_objects/
+        ├── perturbation_assets/
+        ├── scenes/
+        ├── stable_hope_objects/
+        ├── stable_scanned_objects/
+        ├── textures/
+        ├── turbosquid_objects/
+        └── LICENSE_LIBERO
+```
+This is the source layout of the pinned release. If using an already validated
+native LIBERO package layout, the equivalent parent directory is
+`libero/libero/`. The archive does not install Python, MuJoCo, robosuite, model
+checkpoints, or evaluation dependencies. Follow the source release installation
+instructions for those components. Downloading the assets alone does not run
+the benchmark or change the number of evaluation trials.
+## Contents and verification
+The archive contains 1,116 tracked asset files from the source release, plus
+the upstream LIBERO license. All 76 Git LFS assets were expanded and verified
+against their pinned SHA-256 values and sizes. All 181 vendor files match the
+original vendor manifest. For 103 vendor text files, Git had normalized the
+line endings; the original CRLF bytes were restored only after their size and
+SHA-256 matched that source manifest. Other files use canonical Git content.
+`manifests/assets_manifest.json` records the source commit, every source path,
+file size, and SHA-256 checksum. The complete ZIP and every decompressed member
+were checked before publication. This verifies the release payload; it is not
+a new model evaluation or a claim that all policy rollouts succeed.
+The historical asset catalog may mention internal or retired perturbation IDs.
+The authoritative 1,688-task inventory is the pinned source's
+`assets/task_manifest.json`. Public D19 corresponds to internal D20, and public
+D20 corresponds to internal D21. The asset supplement does not rename tasks.
+## Asset provenance and licenses
+The HDF5 dataset's CC BY-NC 4.0 declaration does not replace upstream asset
+licenses. The archive retains the original LIBERO license, vendor licenses,
+and `assets/perturbation_assets/THIRD_PARTY_NOTICES.md` with source attribution.
+The selected asset collection includes MIT, Apache-2.0, CC BY 4.0, and
+LGPL-2.1-only components; consult each retained notice for its scope and any
+additional original dataset terms.

manifests/assets_manifest.json ADDED Viewed

The diff for this file is too large to render. See raw diff

scripts/verify_assets.py ADDED Viewed

	@@ -0,0 +1,57 @@

+"""Verify the LIBERO-Pro asset ZIP or its extracted assets directory."""
+import argparse
+import hashlib
+import json
+import zipfile
+from pathlib import Path, PurePosixPath
+def file_hash(path):
+    checksum = hashlib.sha256()
+    with path.open("rb") as stream:
+        for block in iter(lambda: stream.read(1024 * 1024), b""):
+            checksum.update(block)
+    return checksum.hexdigest()
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--manifest", type=Path, required=True)
+    source = parser.add_mutually_exclusive_group(required=True)
+    source.add_argument("--archive", type=Path)
+    source.add_argument("--assets-dir", type=Path)
+    args = parser.parse_args()
+    manifest = json.loads(args.manifest.read_text(encoding="utf-8"))
+    records = manifest["files"]
+    names = [record["path"] for record in records]
+    if len(records) != manifest["file_count"] or len(set(names)) != len(names):
+        raise ValueError("Manifest count or unique paths mismatch")
+    for name in names:
+        path = PurePosixPath(name)
+        if path.is_absolute() or ".." in path.parts or path.parts[0] != "assets":
+            raise ValueError("Invalid manifest path: " + name)
+    if args.archive:
+        expected = manifest["archive"]
+        if args.archive.stat().st_size != expected["bytes"] or file_hash(args.archive) != expected["sha256"]:
+            raise ValueError("Archive checksum mismatch")
+        with zipfile.ZipFile(args.archive) as archive:
+            if len(archive.infolist()) != len(records) or set(archive.namelist()) != set(names):
+                raise ValueError("Archive file inventory mismatch")
+            for record in records:
+                data = archive.read(record["path"])
+                if len(data) != record["bytes"] or hashlib.sha256(data).hexdigest() != record["sha256"]:
+                    raise ValueError("Asset checksum mismatch: " + record["path"])
+    else:
+        root = args.assets_dir.resolve()
+        for record in records:
+            path = root.joinpath(*PurePosixPath(record["path"]).parts[1:]).resolve()
+            if root not in path.parents or not path.is_file():
+                raise ValueError("Missing or invalid extracted asset: " + record["path"])
+            if path.stat().st_size != record["bytes"] or file_hash(path) != record["sha256"]:
+                raise ValueError("Extracted asset checksum mismatch: " + record["path"])
+    print("Verified {} files from source commit {}".format(len(records), manifest["source_commit"]))
+if __name__ == "__main__":
+    main()