Indy7 ACT spatial-coverage policies
Twelve Action Chunking Transformer (ACT) policies trained on a real Neuromeka Indy7 arm to grasp a vertical PVC-U pipe from a single fixed overhead RGB camera. They are the models behind a study of how the spatial spread of demonstrations affects a visuomotor policy.
- Code, protocol and results: https://github.com/SIRLab-RobotArm/indy7-act-spatial-coverage
- Training data: https://huggingface.co/datasets/SIRLab-HGU/indy7-act-spatial-coverage
- Paper: not published yet; this card will be updated with the DOI
The twelve models
Four training conditions x three seeds (0, 1, 2), each trained for exactly 100,000 optimisation steps with LeRobot's ACT implementation (v0.5.1).
| Condition | Demonstrations | Spread | Success over 120 trials |
|---|---|---|---|
| A | 60 | one position | 12.5% |
| B | 60 | three positions | 61.7% |
| C | 60 | nine positions | 80.0% |
| D | 180 | nine positions | 96.7% |
At positions never demonstrated, condition A succeeded 0 times out of 60, while condition D matched its trained-position rate at 96.7%. The full 480-trial table, the blinding scheme and the analysis plan are in the code repository.
Inputs and outputs
- Input: one 240 x 320 RGB image from the overhead camera, plus a 10-dimensional state (end-effector x, y, z in mm; six joint angles in rad; gripper command).
- Output: a chunk of 4-dimensional actions (end-effector delta x, y, z in mm; gripper command), consumed at 5 Hz.
The policies are tied to this exact camera placement, table, object and robot. They are a reference point for reproducing the study, not a general grasping model.
How they were trained
Trained on the PNG version of the dataset. The published dataset is a video re-encode of it, so retraining from the Hub copy will differ slightly; see the dataset card for the measured pixel difference.
Each condition's episode membership was frozen and checksummed before training, and the evaluation ran under blind model codes.
Each checkpoint keeps the train_config.json it was trained with, including the
absolute paths of the machine it ran on. That is left in place as provenance; it
holds no credentials, and Weights & Biases logging was disabled.
Machine and versions
| Component | Value |
|---|---|
| GPU | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96 GB, driver 595.84 |
| CPU / RAM | AMD Ryzen 7 9800X3D, 16 threads / 123 GB |
| OS | Ubuntu 24.04.4 LTS, kernel 6.8.0 |
| Python | 3.12.13 |
| PyTorch | 2.10.0+cu130, CUDA 13.0 |
| LeRobot | 0.5.1 |
Safety
Running these policies moves a six-axis industrial arm. Keep the workspace clear, keep an emergency stop within reach and never run them unattended. The evaluation harness in the code repository enforces workspace bounds, a per-step limit and a command watchdog; do not bypass them.
Citation
@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
Apache License 2.0. Free to use, modify and redistribute, including commercially, keeping the notices and stating your changes; it also grants the patent rights of the contributors. Full text: https://www.apache.org/licenses/LICENSE-2.0
The training data is released separately under CC BY 4.0; see the dataset card.