DEEL-AI/Hackathon_Team01_v2

This dataset was created using LeRobot.

Dataset Description

This dataset contains the first set of teleoperated demonstrations collected during a two-day hackathon using the LeRobot library and SO-101 robot arms in a leader–follower setup. Each episode shows the follower arm picking one colored cube and placing it onto the matching colored cross inside a 2×2 grid.

Two synchronized RGB cameras were used:

• Top camera: overhead, provides a full 2D view of the workspace (arm, cube, grid).

• Front/low camera: slightly above ground level, facing the arm and grid to capture z-axis cues and arm self-pose.

The background was masked with cardboard panels, but ambient lighting varied throughout the day; this variation is preserved and is useful for robustness studies.

Intended for vision-based imitation learning, multi-view fusion, and tabletop manipulation research.

Use Cases

• Imitation Learning: Behavior cloning from teleop demonstrations.

• Multiview Perception: Fusing top + front perspectives for depth inference without explicit depth sensors.

• Robustness to Lighting: Evaluating policy sensitivity to illumination drift.

• State–Action Alignment: Leveraging synchronized proprioception and images.

• Policy Bootstrapping for curricula: pretrain on single-cube before multi-cube tasks.

Data Collection

Teleoperation & Hardware

• Leader–Follower teleop: human drives a leader arm; follower SO-101 replicates to produce demonstrations.

• Workspace: Tabletop with 2×2 grid; only one cell has a colored cross. One cube is placed in its matching cross per episode.

• Cameras:

  • Front: static overhead.

  • Left: static frontal view emphasizing depth.

• Environment: Cardboard background; illumination changes across time are present in the data.

Episode Protocol

1- Move to pre-grasp and visually localize the target cube.

2- Approach and grasp the cube.

3- Transport and align over the colored cross.

4- Place, release, and return to neutral.

Known Limitations

Lighting drift: Varying brightness/temperature across episodes; apply color constancy, normalization, or photometric augmentation.

Occlusions: Hand/gripper and cube may occlude from the front camera during close approaches.

No depth sensor: Only RGB; consider multi-view fusion or learned depth cues.

Action semantics: Confirm whether actions are delta-pose or joint velocities in each metadata.json.

Early-phase variability: Being the first batch, some episodes may include exploratory motions, hesitations, or failed initial grasps that later recover—useful for learning robustness but consider filtering for clean BC.

Additional Information

• Homepage: deel-ai

• License: apache-2.0

Dataset Structure

meta/info.json:

{
    "codebase_version": "v3.0",
    "robot_type": "so101_follower",
    "total_episodes": 206,
    "total_frames": 84098,
    "total_tasks": 1,
    "chunks_size": 1000,
    "data_files_size_in_mb": 100,
    "video_files_size_in_mb": 500,
    "fps": 30,
    "splits": {
        "train": "0:206"
    },
    "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",
    "features": {
        "action": {
            "dtype": "float32",
            "names": [
                "shoulder_pan.pos",
                "shoulder_lift.pos",
                "elbow_flex.pos",
                "wrist_flex.pos",
                "wrist_roll.pos",
                "gripper.pos"
            ],
            "shape": [
                6
            ]
        },
        "observation.state": {
            "dtype": "float32",
            "names": [
                "shoulder_pan.pos",
                "shoulder_lift.pos",
                "elbow_flex.pos",
                "wrist_flex.pos",
                "wrist_roll.pos",
                "gripper.pos"
            ],
            "shape": [
                6
            ]
        },
        "observation.images.left": {
            "dtype": "video",
            "shape": [
                480,
                640,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 30,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.front": {
            "dtype": "video",
            "shape": [
                480,
                640,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 30,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "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
        }
    }
}