khb2439/capstone-design

capstone design

Robot demonstration datasets collected with an SO101 leader arm and a Piper follower arm. This repository contains finalized dresser*, pickandplace*, and pick&place14 through pick&place33 recording sessions in LeRobot v3.0 format.

Contents

• —42 recording-session datasets, 389 episodes, and 154,762 timesteps.
• —dresser*: 85 episodes and 52,054 timesteps.
• —pickandplace*: 104 episodes and 77,106 timesteps.
• —pick&place14 through pick&place33: 200 episodes and 25,602 timesteps, with English task instructions.
• —Recorded dataset frame rate: 30 FPS.
• —Two RGB camera streams: d405_rgb (424 x 240) and usb_front (320 x 240).
• —Action/state data and MP4 videos are preserved unchanged.
• —New-session task names and episode task descriptions were translated from Korean object labels.
• —Unfinalized PNG recording buffers are excluded.

Each top-level session folder is a separate LeRobot dataset root containing meta/, data/, and videos/. The repository is a collection of datasets; it is not a single merged LeRobot dataset root.

State and action representation

The original 6-dimensional action representation is intentional and is preserved in this release. No action dimensions have been added or removed.

observation.state has 7 components:

[joint_1.pos, joint_2.pos, joint_3.pos, joint_4.pos,
 joint_5.pos, joint_6.pos, gripper.pos]

action has 6 components:

[shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos,
 wrist_flex.pos, wrist_roll.pos, gripper.pos]

The current collection adapter maps the five arm action components to Piper joint_1, joint_2, joint_3, joint_5, and joint_6, respectively. Piper joint_4 is maintained by the adapter and is not a separate recorded action component. State includes its observed position.

The current collection configuration uses normalized joint values in [-100, 100] and gripper values in [0, 100]. Per-session hardware calibration and command-line arguments are not included in the original dataset metadata.

Task annotations

The original dresser* and pickandplace1 through pickandplace12 sessions retain their original test annotations. These older sessions still require semantic labeling for applications using task-specific language instructions.

The 20 sessions pick&place14 through pick&place33 have English instructions in both meta/tasks.parquet and the episode metadata. Their numeric task IDs and action/state data remain unchanged. Korean source labels are retained as provenance in dataset_manifest.json. Representative frames were inspected to name the observed object forms (bottle, can, gum container, tape dispenser).

Download

from huggingface_hub import snapshot_download

dataset_directory = snapshot_download(
    repo_id="khb2439/capstone-design",
    repo_type="dataset",
    local_dir="capstone-design",
)

Use individual session subfolders, such as capstone-design/dresser1, as dataset paths. Configure or convert the data for the selected training loader. For example, this upload preserves LeRobot v3.0; it is not an already converted GR00T LeRobot v2 dataset.

Validation

The original 22-session release had all 129,160 data rows inspected for action/state dimensions, finite values, normalized ranges, and episode/index consistency. Its 44 MP4 files were fully decoded, with 258,320 frames across the two cameras, matching the recorded episode lengths.

For the 20 added sessions, task references were verified across all 200 episodes and 25,602 data rows. Relabeling preserved numeric task IDs, schemas, all non-task metadata columns, and checksums of action/state files and videos. Representative video frames were inspected; full video decoding of these new sessions was not part of this update.

One existing metadata discrepancy is retained unchanged: pickandplace5 contains a stored joint_4 state standard deviation of approximately 0.222004, while direct recomputation gives approximately 0.215845. Recompute statistics when preparing transformed or merged training datasets.

Validation does not certify task success, hardware timestamp synchronization, or model deployment performance.

dataset_manifest.json lists the sessions and SHA-256 checksums of all 294 current data/metadata/video files for integrity verification.