lance_dataset / README.md
czk666's picture
docs: add dataset card metadata
333f514
---
pretty_name: Lance Captured Trajectory Dataset
language:
- en
tags:
- lance
- trajectory
- mano
- robotics
- motion-capture
task_categories:
- robotics
- time-series-forecasting
---
# lance_dataset
This repository stores Lance datasets.
The current dataset is organized as `Dexcanvas.lance/`. It contains Lance data fragments, version manifests, and transaction metadata.
## Dataset Structure
`Dexcanvas.lance` represents a captured trajectory dataset. Each row corresponds to one complete captured trajectory. The field definitions are based on `captured_data_schema.yaml`.
Key conventions:
- `trajectory_metadata.total_frames` defines the sequence length `T`
- All sequence fields must have the same `T` dimension
- All rotations are represented in axis-angle format
- All positions are measured in meters
## MANO Trajectory Fields
MANO-related data is split into metadata fields and time-series fields.
### Metadata
The following fields in `trajectory_metadata` describe the hand layout:
- `hand_names`: variable-length list of hand names
- `mano_hand_shapes`: MANO shape parameters for each hand
Index mapping:
- `hand_names[i]` corresponds to `hands[i]`
- `mano_hand_shapes[i]` corresponds to `hands[i]`
### Time-Series Data
`hands[]` is a variable-length array. Each element represents one hand over the full trajectory.
- `mano_global_pos` `(T, 3)`
Global hand position in meters.
- `mano_global_rot_aa` `(T, 3)`
Global hand rotation in axis-angle format.
- `mano_hand_pose` `(T, 48)`
MANO hand pose parameters.
- 45 dimensions for 15 joint rotations
- 3 dimensions for the wrist
- `mano_joint_pos` `(T, 21, 3)`
3D positions of 21 MANO hand joints.
- `urdf_dof` `(T, 26)`
URDF degree-of-freedom representation.
- 20 finger DOFs
- 6 wrist DOFs
In short:
- `mano_global_pos` and `mano_global_rot_aa` describe the global hand pose
- `mano_hand_pose` describes MANO finger and wrist pose parameters
- `mano_joint_pos` provides the derived 3D hand joint positions
- `urdf_dof` provides a robot/control-oriented representation
## Object Fields
Object-related data is stored in `trajectory_metadata.object_names` and `objects[]`.
### Metadata
- `object_names`: variable-length list of object names
Index mapping:
- `object_names[i]` corresponds to `objects[i]`
### Time-Series Data
`objects[]` is a variable-length array. Each element represents one object over the full trajectory.
- `rot_aa` `(T, 3)`
Object rotation in axis-angle format.
- `pos` `(T, 3)`
Object position in meters.
Therefore, `objects[i]` stores the full pose trajectory of the `i`-th object across all `T` frames.
## Other Common Fields
- `index.uuid`: unique trajectory identifier
- `index.seed_uuid`: same as `uuid` for captured data; source trajectory ID for generated data
- `index.file_uuid`: source `session_data_files.file_uuid`
- `index.capMachine`: capture machine identifier
- `index.operator`: operator who captured or generated the data
- `index.scene`: scene identifier
- `index.is_generated`: whether the row is generated data
- `timestamp`: per-frame timestamps
- `atomic_actions`: atomic action intervals, such as grasp or release
## Directory Layout
- `Dexcanvas.lance/`: Lance dataset root
- `Dexcanvas.lance/data/`: physical Lance data fragments
- `Dexcanvas.lance/_versions/`: dataset version manifests
- `Dexcanvas.lance/_transactions/`: transaction records
- `Dexcanvas.lance/_latest.manifest`: latest dataset manifest
- `captured_data_schema.yaml`: YAML schema converted from the upstream JSONC schema
## Notes
When adding new fields, keep them consistent with `captured_data_schema.yaml`, especially:
- The shared `T` dimension for sequence fields
- The index mapping between `hands[]` and `trajectory_metadata.hand_names`
- The index mapping between `objects[]` and `trajectory_metadata.object_names`
- Axis-angle rotation representation