# InstructMove Dataset

## Version

**Version 1.0**

InstructMove is a language-conditioned robot manipulation dataset in
[LeRobot](https://github.com/huggingface/lerobot) format. It contains
instruction-following trajectories for object picking, spatially grounded
picking, and object placement. The current release includes data collected
with PiperX dual-arm and Franka robots.

## Dataset Contents

| Dataset | Task | Robot | Path | Episodes |
| --- | --- | --- | --- | ---: |
| Pick by attribute | Pick an object specified by visual attributes such as color, shape, or appearance. | PiperX dual-arm | `lerobot_dataset/pick_attribute/piperx_0527` | 5,601 |
| Pick by category | Pick an object specified by its semantic category and visual description. | PiperX dual-arm | `lerobot_dataset/pick_category/piperx_0527` | 8,904 |
| Place A to B | Pick a specified object and place it in or on a target container. | PiperX dual-arm | `lerobot_dataset/place_a2b/piperx_0527` | 10,005 |
| Place A to B | Pick a specified object and place it in or on a target container. | Franka | `lerobot_dataset/place_a2b/franka_0612` | 4,038 |
| Spatial pick | Pick an object identified through a spatial relation to another object, such as left of, right of, near, or far from. | Franka | `lerobot_dataset/spatial_pick/franka_0612` | 3,618 |
| **Total** |  |  | `lerobot_dataset` | **32,166** |

## Dataset Structure

Each row in the table corresponds to an independent LeRobot dataset. Its
directory contains trajectory data, videos, and metadata, including episode
and task annotations. Episode counts are taken from each dataset's
`meta/info.json` and verified against `meta/episodes.jsonl`.