metadata
task_categories:
- robotics
tags:
- LeRobot
- d3il
- imitation-learning
Combined D3IL dataset (LeRobot v2.1)
This repository combines all eight D3IL demonstration datasets into one task-conditioned LeRobot v2.1 dataset.
observation.state is zero-padded to 22 values and action to
8 values. state_is_pad and action_is_pad identify padding.
source_episode_index preserves the episode number within the source dataset,
and source_fps preserves its original nominal FPS. The repository FPS is
29, the rate shared by seven of the eight source datasets; no source
frames were dropped or interpolated.
Camera features are not included because four source tasks do not have camera data and LeRobot v2.1 does not support optional per-task features.
| Source | Task instruction | State dim | Action dim | Source FPS |
|---|---|---|---|---|
d3il_avoiding |
Move the end effector to the goal while avoiding the obstacles. | 4 | 2 | 29 |
d3il_aligning |
Align the push box with the target box. | 20 | 3 | 29 |
d3il_inserting |
Insert the three boxes into their target gates. | 13 | 2 | 29 |
d3il_pushing |
Push the red and green boxes to their target positions. | 10 | 2 | 29 |
d3il_sorting_2 |
Sort the two boxes by color. | 10 | 2 | 29 |
d3il_sorting_4 |
Sort the four boxes by color. | 16 | 2 | 29 |
d3il_sorting_6 |
Sort the six boxes by color. | 22 | 2 | 29 |
d3il_stacking |
Stack the red, green, and blue boxes. | 20 | 8 | 33 |