willlyb/snack_transfer

This dataset was created using LeRobot.

This is a reconstructed dataset. Original recording durations are preserved with held images/poses across missing intervals. Wrist-camera timing is approximate in all episodes; 41.46 seconds have unsupported poses, marked by is_lost. It is not verified one-to-one synchronized training data. See recovery details and masks.

Dataset Structure

Two video observations (observation.images.left_cam0, observation.images.left_cam1), eight action values (position, rotation vector, and two gripper angles), and is_lost, using the same feature schema as ParallaxData/sliding_door_clean.

meta/info.json:

{
    "codebase_version": "v3.0",
    "fps": 50,
    "features": {
        "observation.images.left_cam0": {
            "dtype": "video",
            "shape": [
                3,
                720,
                960
            ],
            "names": [
                "channels",
                "height",
                "width"
            ],
            "info": {
                "video.height": 720,
                "video.width": 960,
                "video.codec": "h264",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 50,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.left_cam1": {
            "dtype": "video",
            "shape": [
                3,
                720,
                960
            ],
            "names": [
                "channels",
                "height",
                "width"
            ],
            "info": {
                "video.height": 720,
                "video.width": 960,
                "video.codec": "h264",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 50,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "action": {
            "dtype": "float32",
            "shape": [
                8
            ],
            "names": [
                "left_x",
                "left_y",
                "left_z",
                "left_ax",
                "left_ay",
                "left_az",
                "left_proximal",
                "left_distal"
            ]
        },
        "is_lost": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": [
                "left_is_lost"
            ]
        },
        "timestamp": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": null
        },
        "frame_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "episode_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "task_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        }
    },
    "total_episodes": 10,
    "total_frames": 7324,
    "total_tasks": 1,
    "chunks_size": 1000,
    "data_files_size_in_mb": 100,
    "video_files_size_in_mb": 200,
    "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
    "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
    "robot_type": "grabette",
    "splits": {
        "train": "0:10"
    }
}

Recovery Metadata

Additional validity flags are stored separately in meta/recovery_masks.parquet so they do not add charts to the standard visualizer. Join them by index or by episode_index and frame_index; standard training code does not load them automatically. Per-frame provenance and validation results are also retained.

Citation

No associated paper.