| --- |
| pretty_name: Tower-SimData |
| license: cc-by-4.0 |
| language: |
| - en |
| tags: |
| - robotics |
| - bimanual-manipulation |
| - imitation-learning |
| - jenga |
| - simulation |
| - hdf5 |
| --- |
| |
| # Tower-SimData |
|
|
| Tower-SimData is a simulated demonstration dataset for research on bimanual Jenga manipulation. Each episode is stored as an independent HDF5 file containing bimanual robot states and commands, end-effector poses, four JPEG camera streams, and simulation task states. |
|
|
| Dataset repository: <https://huggingface.co/datasets/tower-benchmark/TOWER-SimData> |
|
|
| ## Dataset Size |
|
|
| | Task | Name | Public folder | Episodes | Frames | Size | |
| |---|---|---|---:|---:|---:| |
| | Task1-2 | Tower Transfer, 2 layers | `task1-2_tower_transfer_2` | 56 | 196,031 | 45.99 GiB | |
| | Task1-3 | Tower Transfer, 3 layers | `task1-3_tower_transfer_3` | 55 | 294,441 | 67.08 GiB | |
| | Task2-2 | Cross Tower Transfer, 2 layers | `task2-2_cross_tower_transfer_2` | 55 | 100,086 | 22.26 GiB | |
| | Task2-3 | Cross Tower Transfer, 3 layers | `task2-3_cross_tower_transfer_3` | 55 | 155,176 | 35.19 GiB | |
| | Task3-2 | Vertical Stacking, 2 layers | `task3-2_vertical_stacking_2` | 55 | 97,124 | 22.19 GiB | |
| | Task3-3 | Vertical Stacking, 3 layers | `task3-3_vertical_stacking_3` | 55 | 140,900 | 32.55 GiB | |
| | Task4-2 | Full Tower Pushdown, 2 layers | `task4-2_full_tower_pushdown_2` | 50 | 177,203 | 40.65 GiB | |
| | Task4-3 | Full Tower Pushdown, 3 layers | `task4-3_full_tower_pushdown_3` | 50 | 258,895 | 60.68 GiB | |
| | Task5-2 | Tower Insertion, 2 layers | `task5-2_tower_insertion_2` | 50 | 88,549 | 19.95 GiB | |
| | Task5-3 | Tower Insertion, 3 layers | `task5-3_tower_insertion_3` | 50 | 240,175 | 57.13 GiB | |
| | **Total** | | | **531** | **1,748,580** | **403.68 GiB** | |
|
|
| The public release does not define an official training or test split. All 531 episodes are released together. The Task4/Task5 addition contains 50 episodes for each task/layer, selected from the source training pools; this does not establish a public benchmark split. Users should create and document splits that match their experimental protocol. |
|
|
| ## Repository Layout |
|
|
| ```text |
| data/ |
| ├── task1-2_tower_transfer_2/ |
| ├── task1-3_tower_transfer_3/ |
| ├── task2-2_cross_tower_transfer_2/ |
| ├── task2-3_cross_tower_transfer_3/ |
| ├── task3-2_vertical_stacking_2/ |
| ├── task3-3_vertical_stacking_3/ |
| ├── task4-2_full_tower_pushdown_2/ |
| ├── task4-3_full_tower_pushdown_3/ |
| ├── task5-2_tower_insertion_2/ |
| └── task5-3_tower_insertion_3/ |
| previews/ |
| MANIFEST.jsonl |
| CHECKSUMS.sha256 |
| schema/tower_simdata_hdf5.json |
| ``` |
|
|
| `MANIFEST.jsonl` records the relative path, frame count, file size, SHA256 digest, collection date, and schema variant of every episode. `CHECKSUMS.sha256` can be used to verify download integrity. |
|
|
| ## Task Definitions and Naming |
|
|
| `N` denotes the task index. `L` and `X` denote the number of layers in the block structure. Public folders follow `taskN-L_task_name`; for example, `task1-2_tower_transfer_2` contains the two-layer variant of Task1. |
|
|
| | Task | Task name | Folder naming pattern | |
| |---|---|---| |
| | Task1 | Tower Transfer | `task1-X_tower_transfer_X` | |
| | Task2 | Cross Tower Transfer | `task2-X_cross_tower_transfer_X` | |
| | Task3 | Vertical Stacking | `task3-X_vertical_stacking_X` | |
| | Task4 | Full Tower Pushdown | `task4-X_full_tower_pushdown_X` | |
| | Task5 | Tower Insertion | `task5-X_tower_insertion_X` | |
|
|
| ## Core HDF5 Fields |
|
|
| | Field | Shape | Description | |
| |---|---|---| |
| | `observations/qpos` | `[T,16]` | Recorded bimanual state: seven arm joints and one gripper value per arm | |
| | `leader/qpos` | `[T,16]` | Absolute bimanual control commands: seven arm joints and one gripper value per arm | |
| | `observations/gripper_status` | `[T,8]` | Extended gripper state, four values per gripper | |
| | `observations/eef` | `[T,14]` | Bimanual end-effector poses as `xyz + quaternion(xyzw)` | |
| | `leader/eef` | `[T,14]` | Command-side end-effector poses | |
| | `observations/lifter/position_mm` | `[T,2]` | Left and back lifter position readings | |
| | `observations/lifter/physical_height_mm` | `[T,2]` | Physical heights of the left and back lifters | |
| | `observations/lifter/is_moving` | `[T,2]` | Whether each lifter is moving | |
| | `observations/lifter/valid` | `[T,2]` | Whether each lifter observation is valid | |
| | `lifter/action_mm` | `[T,2]` | Absolute commands for the left and back lifters | |
| | `lifter/active_target_mm` | `[T,2]` | Active target of each lifter | |
| | `lifter/command_delta_mm` | `[T,2]` | Difference between consecutive lifter commands | |
| | `observations/images/*` | `[T]` | Four variable-length JPEG camera streams | |
| | `sim/time` | `[T]` | Strictly increasing simulation time | |
| | `sim/action_valid` | `[T]` | Action-valid flag; the final row is `false` | |
| | `sim/terminal` | `[T]` | Episode termination flag | |
| | `sim/success` | `[T]` | Task success state | |
|
|
| The fixed arm order is `left, back`. The four cameras are `left_back_camera`, `right_front_camera`, `left_wrist_camera`, and `back_wrist_camera`. The source sampling rate is 30 FPS. |
|
|
| All tasks use the unified `tower_simdata_hdf5` schema. The 37 required fields and optional fields are defined in `schema/tower_simdata_hdf5.json`. `keyframes` is an optional general-purpose field. Task2, Task3 and Task5 episodes may additionally contain `sim/finger_pad_centers_m`. Task4/Task5 episodes also include optional simulator diagnostics: `sim/controller_qpos`, `sim/gripper_joint_positions_m`, and `sim/gripper_requested_width_m`. These do not replace the standard state or command fields; existing readers can ignore them. |
|
|
| ### Camera orientation and compatibility |
|
|
| Stored JPEGs are already oriented for use: decode them directly, without an extra rotation or flip. For the v1.1.0 Task4/Task5 addition, the two external cameras were losslessly rotated by 180 degrees to match the original public release. Wrist camera JPEGs and all non-image dataset values are unchanged. `camera_transforms` and per-camera `orientation_transform` describe the transform already applied, not a transform to apply during loading. Simulator configuration JSON preserves the historical capture settings. |
|
|
| The HDF5 core schema, 30 FPS timeline, units, arm order and action alignment are unchanged. This is a raw HDF5 release; a model-specific conversion and normalization step is still required for training. Existing readers should discover the four new task folders and retain their existing preprocessing conventions. Recompute normalization for the actual training split rather than treating this addition as a pre-normalized dataset. |
|
|
| ### 18-Dimensional Action Convention |
|
|
| The public HDF5 files store the 18-dimensional action in two synchronized fields instead of duplicating it as an additional `action[T,18]` dataset: |
|
|
| ```text |
| action[18] = concat( |
| leader/qpos[16], |
| lifter/action_mm[2] |
| ) |
| |
| leader/qpos[16] = [ |
| left_J1 ... left_J7, left_gripper, |
| back_J1 ... back_J7, back_gripper |
| ] |
| |
| lifter/action_mm[2] = [left_lifter, back_lifter] |
| ``` |
|
|
| J1–J7 are measured in radians, gripper values are continuous absolute opening commands, and lifter commands are measured in millimeters. These quantities must be concatenated in the order above and must not be interpreted as sharing one physical unit. |
|
|
| Measured gripper observations retain simulator excursions around zero (the Task4/Task5 minimum is approximately -1.32 mm); command widths remain nonnegative. The release does not clip these raw values. Retain the existing gripper tolerance and canonicalization step in training preprocessing. |
|
|
| ## Camera Previews |
|
|
| Each preview uses one representative episode. The first row shows the initial frame and the second row shows the final frame. Within each row, the four camera views are shown in a fixed order. |
|
|
| ### Task1-2 · `task1-2_tower_transfer_2` |
| |
|  |
| |
| ### Task1-3 · `task1-3_tower_transfer_3` |
|
|
|  |
|
|
| ### Task2-2 · `task2-2_cross_tower_transfer_2` |
|
|
|  |
|
|
| ### Task2-3 · `task2-3_cross_tower_transfer_3` |
|
|
|  |
|
|
| ### Task3-2 · `task3-2_vertical_stacking_2` |
| |
|  |
| |
| ### Task3-3 · `task3-3_vertical_stacking_3` |
|
|
|  |
|
|
| ### Task4-2 · `task4-2_full_tower_pushdown_2` |
|
|
|  |
|
|
| ### Task4-3 · `task4-3_full_tower_pushdown_3` |
|
|
|  |
|
|
| ### Task5-2 · `task5-2_tower_insertion_2` |
| |
|  |
| |
| ### Task5-3 · `task5-3_tower_insertion_3` |
|
|
|  |
|
|
| ## Reading and Verification |
|
|
| ```bash |
| pip install h5py |
| |
| python - <<'PY' |
| import h5py |
| |
| with h5py.File("data/task1-3_tower_transfer_3/episode_xxx.hdf5", "r") as episode: |
| print(episode["observations/qpos"].shape) |
| print(list(episode["observations/images"])) |
| PY |
| |
| sha256sum --check CHECKSUMS.sha256 |
| ``` |
|
|
| The dataset preserves the original 30 FPS timeline. No model-specific action transformation or normalization has been applied. |
|
|
| ### Downloading only the Task4/Task5 addition |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| |
| snapshot_download( |
| repo_id="tower-benchmark/TOWER-SimData", |
| repo_type="dataset", |
| local_dir="TOWER-SimData", |
| allow_patterns=["data/task4-*/*.hdf5", "data/task5-*/*.hdf5", |
| "MANIFEST.jsonl", "CHECKSUMS.sha256", "schema/*", |
| "README.md", "LICENSE", "CITATION.cff"], |
| ) |
| ``` |
|
|
| The root manifest and checksum file cover the entire repository. When downloading only selected tasks, verify only those entries whose files were downloaded; missing files from other tasks are expected. |
|
|
| ## License and Citation |
|
|
| Tower-SimData is released under the [Creative Commons Attribution 4.0 International License](LICENSE). If you use this dataset, please cite the metadata in [`CITATION.cff`](CITATION.cff). |
|
|
