docs: richer dataset README (conversion details, v3 structure, citation)

README.md

@@ -2,37 +2,87 @@
 license: other
 task_categories:
 - robotics
+language:
+- en
 tags:
-- lerobot
-- rlbench
+- simulation
+- single-arm
 - robot-manipulation
 - visuomotor-policy
+- LeRobot
+- v3
+- end-effector-pose
+- rlbench
+size_categories:
+- 100K<n<1M
+pretty_name: "RLBench-18 Train (LeRobot v3 EE)"
+viewer: false
 ---
 
 # RLBench-18 v3 EEF — Train Split (LeRobot v3)
 
-RLBench 18-task PerAct dataset converted to LeRobot v3 with lossless H.264
-video (CRF=0, yuv444p), 20 Hz, single-arm Franka Panda.
+This is a **LeRobot v3** format conversion of the 18-task [RLBench](https://github.com/stepjam/RLBench) training split (PerAct demonstration format), reorganized into a canonical **8D end-effector (EE) pose** representation for multi-task visuomotor policy training. Lossless H.264 video (CRF=0, yuv444p), 20 Hz, single-arm Franka Panda.
+
+## Original Dataset
+
+**RLBench: The Robot Learning Benchmark & Learning Environment**
+
+> James, S., Ma, Z., Rovick Arrojo, D., Davison, A. J. *RLBench: The Robot Learning Benchmark & Learning Environment*. arXiv:1909.12271, IEEE Robotics and Automation Letters, 2020.
+
+- **Paper**: https://arxiv.org/abs/1909.12271
+- **GitHub**: https://github.com/stepjam/RLBench
+- **Demonstration format**: PerAct (Shridhar et al., CoRL 2022) — https://github.com/peract/peract
+
+RLBench is a large-scale simulated manipulation benchmark built on CoppeliaSim, with procedurally generated task variations. This conversion uses the 18-task subset popularized by PerAct, with 100 demonstrations per task in the train split.
+
+## License
+
+The original RLBench code and assets are released under the MIT license; the PerAct demonstrations inherit their respective terms. This conversion is provided under `license: other` — please consult the upstream repositories for details.
+
+## Conversion Details
 
-## Layout
+### What we changed
 
-Each task is a self-contained LeRobot dataset in its own subdirectory
-(`close_jar/`, `insert_onto_square_peg/`, ...) with `meta/`, `data/`, and
-`videos/`. 100 episodes per task, ~374K frames total.
+1. **Canonical EE Pose Representation**: Per-frame proprioception is unified into a single **8D EE pose vector** — position xyz (metres) + quaternion in scalar-first **WXYZ** order + gripper.
 
-## Observation and action convention
+2. **Next-Observation-Absolute Actions**: Actions are stored as **8D absolute EE pose commands**. Row `t` stores the pose that the *next* observation row `t+1` attains (next-observation-absolute alignment). Episodes contain exactly `observations - 1` action rows, so no terminal placeholder rows exist.
 
-- `observation.state`: 8D end-effector pose — position xyz (metres) +
-  quaternion in scalar-first WXYZ order + gripper.
-- `action`: 8D absolute end-effector pose command. Row `t` stores the pose
-  that the *next* observation row `t+1` attains (next-observation-absolute
-  alignment). Episodes contain exactly `observations - 1` action rows, so no
-  terminal placeholder rows exist.
+3. **LeRobot v3 Format**: Converted to the LeRobot v3 dataset layout with sharded lossless H.264 video (CRF=0, yuv444p) and Parquet-based frame data at 20 Hz.
+
+### What we preserved
+
+- Both camera views: `observation.images.head` (front view) and `observation.images.wrist_right` (wrist view), 128x128 RGB
+- Episode structure, task labels, and task-level variation counts
+- Frame-level timestamps
+
+### Observation and action convention
+
+- `observation.state`: 8D end-effector pose — `[x, y, z, qw, qx, qy, qz, gripper]`.
+- `action`: 8D absolute end-effector pose command with the same layout; row `t` is the pose attained by observation row `t+1`.
 - Gripper is binary: `0` = closed, `1` = open.
-- Cameras: `observation.images.head` (front view) and
-  `observation.images.wrist_right` (wrist view), 256x256 RGB.
 
-## Task list
+### Layout
+
+Each task is a self-contained LeRobot v3 dataset in its own subdirectory with its own `meta/`, `data/`, and `videos/`:
+
+```
+rlbench_v3_eef/
+├── close_jar/
+│   ├── data/chunk-*/file-*.parquet             # Frame data (8D state + 8D action)
+│   ├── videos/
+│   │   ├── observation.images.head/chunk-*/file-*.mp4
+│   │   └── observation.images.wrist_right/chunk-*/file-*.mp4
+│   └── meta/
+│       ├── info.json                           # Dataset metadata
+│       ├── tasks.parquet                       # Task/variation vocabulary
+│       ├── episodes/chunk-*/file-*.parquet     # Episode index
+│       └── stats.json                          # Per-feature dataset statistics
+├── insert_onto_square_peg/
+│   └── ... (18 task directories total)
+└── norm_stats.json                             # Split-level post-pipeline stats (schema v3)
+```
+
+### Task list
 
 close_jar, insert_onto_square_peg, light_bulb_in, meat_off_grill, open_drawer,
 place_cups, place_shape_in_shape_sorter, place_wine_at_rack_location,
@@ -40,21 +90,44 @@ push_buttons, put_groceries_in_cupboard, put_item_in_drawer, put_money_in_safe,
 reach_and_drag, slide_block_to_color_target, stack_blocks, stack_cups,
 sweep_to_dustpan_of_size, turn_tap
 
-## Acknowledgments
+### Statistics
 
-Data: PerAct format. Simulator: RLBench.
+| Metric | Value |
+|---|---|
+| Total Tasks | 18 |
+| Total Episodes | 1,800 (100 per task) |
+| Total Frames | 373,767 |
+| FPS | 20 |
+| Robot Type | Franka Panda |
+| State Dim | 8 |
+| Action Dim | 8 |
+
+## Usage
+
+```python
+from lerobot.datasets import LeRobotDataset
+
+# Each task is an independent LeRobot dataset under a subdirectory
+dataset = LeRobotDataset("GT-111/rlbench_v3_eef", root=".../close_jar")
+```
+
+## Citation
 
 ```bibtex
-@inproceedings{shridhar2022peract,
-  title={Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation},
-  author={Shridhar, Mohit and Manuelli, Lucas and Fox, Dieter},
-  booktitle={CoRL},
-  year={2022}
-}
 @article{james2019rlbench,
   title={RLBench: The Robot Learning Benchmark & Learning Environment},
   author={James, Stephen and Ma, Zicong and Rovick Arrojo, David and Davison, Andrew J},
   journal={IEEE Robotics and Automation Letters},
   year={2020}
 }
+@inproceedings{shridhar2022peract,
+  title={Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation},
+  author={Shridhar, Mohit and Manuelli, Lucas and Fox, Dieter},
+  booktitle={CoRL},
+  year={2022}
+}
 ```
+
+## Version History
+
+- **v3.0** (current): LeRobot v3 conversion with canonical 8D EE pose layout and next-observation-absolute actions