Datasets:

SIRLab-HGU
/

indy7-act-spatial-coverage

---
license: cc-by-4.0
task_categories:
  - robotics
tags:
  - LeRobot
  - imitation-learning
  - action-chunking-transformer
  - manipulation
  - real-robot
configs:
  - config_name: default
    data_files: data/*/*.parquet
---
# Indy7 ACT spatial-coverage dataset

880 episodes (176,837 frames) of a Neuromeka Indy7 arm with a Mand.ro Mark7 hand
grasping a vertical PVC-U pipe, seen by a single fixed overhead RGB camera. The
data backs a study of one question: does it matter more *where* demonstrations
are collected, or *how many* there are?

- Code, protocol and results: https://github.com/SIRLab-RobotArm/indy7-act-spatial-coverage
- Trained policies: https://huggingface.co/SIRLab-HGU/indy7-act-spatial-coverage-models
- Paper: not published yet; this card will be updated with the DOI

## What is in it

| Field | Shape | Meaning |
| --- | --- | --- |
| `observation.images.overhead` | 240 x 320 x 3, video | Fixed overhead camera, colour only, no depth |
| `observation.state` | 10 | End-effector x, y, z in mm; six joint angles in rad; gripper command |
| `action` | 4 | End-effector delta x, y, z in mm to the next control step; gripper command |

- 880 episodes, 176,837 frames, 5 Hz, one task: *grasp and lift the cylinder*.
- Episodes come from 220 human teleoperated recordings at 20 Hz. Each recording
  is split into four 5 Hz episodes by phase, so 220 x 4 = 880.
- Recordings were made at nine marked table positions (G1-G9), 20 per position
  plus 40 extra at the centre.

## Loading

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("SIRLab-HGU/indy7-act-spatial-coverage")
print(dataset.meta.total_episodes, dataset.meta.total_frames)
sample = dataset[0]
```

## The four training conditions

The study compares demonstration *spread* against demonstration *count*. The
episode membership of each condition was frozen and checksummed before any model
was trained; the manifests live in `experiment/manifests/` in the code
repository and refer to the episode indices used here.

| Condition | Demonstrations | Spread |
| --- | ---: | --- |
| A | 60 | one position only |
| B | 60 | three positions |
| C | 60 | nine positions |
| D | 180 | nine positions |

## How this file set was produced

The released videos are a re-encode of the image dataset the published policies
were trained on, not a fresh conversion of the raw recordings. Episode indices,
episode lengths, states, actions, timestamps and task indices are identical to
the published artefacts; only the pixels differ, because PNG frames were encoded
as AV1 video (`libsvtav1`, CRF 30, GOP 2, `yuv420p`) with LeRobot's own
`convert_image_to_video_dataset`.

Measured difference over 500 randomly sampled frames:

| Metric | Value |
| --- | ---: |
| Max absolute pixel difference | 93 |
| Mean absolute pixel difference | 1.8269 |
| PSNR, mean | 40.36 dB |
| PSNR, worst frame | 38.29 dB |

Part of that difference is the RGB to `yuv420p` conversion, which is lossy even
with a lossless codec. Policies in the accompanying release were trained on the
PNG version; expect small differences if you retrain from these videos.

The raw 20 Hz HDF5 recordings (90 GB) are not published here. Ask the authors if
you need them.

## Collection setup

| Part | What was used |
| --- | --- |
| Arm | Neuromeka Indy7 |
| Hand | Mand.ro Mark7 |
| Camera | Intel RealSense D435, fixed above the table, colour only |
| Object | 25 cm rigid PVC-U pipe, 2.5 cm outer diameter |
| Teleoperation | Xbox controller |

## Limitations

- One object, one camera pose, one lighting setup, one laboratory.
- Demonstrations were collected position by position in blocks, so time of day,
  operator fatigue and lighting drift are entangled with position.
- Only successful demonstrations were kept, so the data contains no failures.
- Frames show the robot and the table; no people are intentionally recorded.

## Citation

```bibtex
@software{kim_indy7_act_spatial_coverage_2026,
  author = {Kim, Yuyoung and Kim, Dogyung and Yoo, Jaehyeok and
            Kim, Dongyun and Kim, Dowook and Lee, WonHyong},
  title  = {Spatial coverage versus sample count in visuomotor imitation
            learning: an RGB-only ACT study on a real Indy7 arm},
  year   = {2026},
  url    = {https://github.com/SIRLab-RobotArm/indy7-act-spatial-coverage}
}
```

The paper citation will be added here once it is published.

## License

Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and
adapt this dataset, including commercially, as long as you credit the authors and
say whether you changed it. Full text: https://creativecommons.org/licenses/by/4.0/

The Mark7 hand geometry and the Neuromeka Indy7 driver are not part of this
dataset. See the code repository's `NOTICE.md` for third-party components.

	---
	license: cc-by-4.0
	task_categories:
	- robotics
	tags:
	- LeRobot
	- imitation-learning
	- action-chunking-transformer
	- manipulation
	- real-robot
	configs:
	- config_name: default
	data_files: data//.parquet
	---

	# Indy7 ACT spatial-coverage dataset

	880 episodes (176,837 frames) of a Neuromeka Indy7 arm with a Mand.ro Mark7 hand
	grasping a vertical PVC-U pipe, seen by a single fixed overhead RGB camera. The
	data backs a study of one question: does it matter more where demonstrations
	are collected, or how many there are?

	- Code, protocol and results: https://github.com/SIRLab-RobotArm/indy7-act-spatial-coverage
	- Trained policies: https://huggingface.co/SIRLab-HGU/indy7-act-spatial-coverage-models
	- Paper: not published yet; this card will be updated with the DOI

	## What is in it

	\| Field \| Shape \| Meaning \|
	\| --- \| --- \| --- \|
	\| `observation.images.overhead` \| 240 x 320 x 3, video \| Fixed overhead camera, colour only, no depth \|
	\| `observation.state` \| 10 \| End-effector x, y, z in mm; six joint angles in rad; gripper command \|
	\| `action` \| 4 \| End-effector delta x, y, z in mm to the next control step; gripper command \|

	- 880 episodes, 176,837 frames, 5 Hz, one task: grasp and lift the cylinder.
	- Episodes come from 220 human teleoperated recordings at 20 Hz. Each recording
	is split into four 5 Hz episodes by phase, so 220 x 4 = 880.
	- Recordings were made at nine marked table positions (G1-G9), 20 per position
	plus 40 extra at the centre.

	## Loading

	```python
	from lerobot.datasets.lerobot_dataset import LeRobotDataset

	dataset = LeRobotDataset("SIRLab-HGU/indy7-act-spatial-coverage")
	print(dataset.meta.total_episodes, dataset.meta.total_frames)
	sample = dataset[0]
	```

	## The four training conditions

	The study compares demonstration spread against demonstration count. The
	episode membership of each condition was frozen and checksummed before any model
	was trained; the manifests live in `experiment/manifests/` in the code
	repository and refer to the episode indices used here.

	\| Condition \| Demonstrations \| Spread \|
	\| --- \| ---: \| --- \|
	\| A \| 60 \| one position only \|
	\| B \| 60 \| three positions \|
	\| C \| 60 \| nine positions \|
	\| D \| 180 \| nine positions \|

	## How this file set was produced

	The released videos are a re-encode of the image dataset the published policies
	were trained on, not a fresh conversion of the raw recordings. Episode indices,
	episode lengths, states, actions, timestamps and task indices are identical to
	the published artefacts; only the pixels differ, because PNG frames were encoded
	as AV1 video (`libsvtav1`, CRF 30, GOP 2, `yuv420p`) with LeRobot's own
	`convert_image_to_video_dataset`.

	Measured difference over 500 randomly sampled frames:

	\| Metric \| Value \|
	\| --- \| ---: \|
	\| Max absolute pixel difference \| 93 \|
	\| Mean absolute pixel difference \| 1.8269 \|
	\| PSNR, mean \| 40.36 dB \|
	\| PSNR, worst frame \| 38.29 dB \|

	Part of that difference is the RGB to `yuv420p` conversion, which is lossy even
	with a lossless codec. Policies in the accompanying release were trained on the
	PNG version; expect small differences if you retrain from these videos.

	The raw 20 Hz HDF5 recordings (90 GB) are not published here. Ask the authors if
	you need them.

	## Collection setup

	\| Part \| What was used \|
	\| --- \| --- \|
	\| Arm \| Neuromeka Indy7 \|
	\| Hand \| Mand.ro Mark7 \|
	\| Camera \| Intel RealSense D435, fixed above the table, colour only \|
	\| Object \| 25 cm rigid PVC-U pipe, 2.5 cm outer diameter \|
	\| Teleoperation \| Xbox controller \|

	## Limitations

	- One object, one camera pose, one lighting setup, one laboratory.
	- Demonstrations were collected position by position in blocks, so time of day,
	operator fatigue and lighting drift are entangled with position.
	- Only successful demonstrations were kept, so the data contains no failures.
	- Frames show the robot and the table; no people are intentionally recorded.

	## Citation

	```bibtex
	@software{kim_indy7_act_spatial_coverage_2026,
	author = {Kim, Yuyoung and Kim, Dogyung and Yoo, Jaehyeok and
	Kim, Dongyun and Kim, Dowook and Lee, WonHyong},
	title = {Spatial coverage versus sample count in visuomotor imitation
	learning: an RGB-only ACT study on a real Indy7 arm},
	year = {2026},
	url = {https://github.com/SIRLab-RobotArm/indy7-act-spatial-coverage}
	}
	```

	The paper citation will be added here once it is published.

	## License

	Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and
	adapt this dataset, including commercially, as long as you credit the authors and
	say whether you changed it. Full text: https://creativecommons.org/licenses/by/4.0/

	The Mark7 hand geometry and the Neuromeka Indy7 driver are not part of this
	dataset. See the code repository's `NOTICE.md` for third-party components.