SN0429/so101_pick_cube_test_20260823_40ep_urdf

This dataset was created using LeRobot.

What this is

The URDF-aligned repair of SN0429/so101_pick_cube_test_20260823_40ep, for 3D visualisation and Isaac Sim.

LeRobot's degree zero means "the pose held during calibration"; the SO-101 URDF's zero is a different configuration. On elbow_flex the two are about 97 degrees apart and opposite in sign, so a recording in LeRobot's convention renders in the wrong posture. This copy maps the angles into the URDF's frame.

Joint	Correction	Result
elbow_flex	unwrapped, then urdf = -lerobot - 96.62	-180.0..0.2 -> -96.8..84.0, inside the URDF limit +-96.8
wrist_roll	shifted -155.25 deg into the corrected calibration frame	136.6..159.7 -> -18.7..4.4
the other four	untouched

Validated against a photo of the arm at rest. At frame 0 this mapping puts the arm in a tall vertical column - tip 0.047 m horizontally from the base and 0.520 m up, an 11:1 aspect - which is what the real SO-101 looks like in its rest pose. The two alternatives considered (leaving the angles untransformed, or a +90 deg offset without the sign flip) put the arm horizontal and folded below the base respectively, and neither matches the photo.

Known limitation: about 20% of frames put the gripper below the base plane. The arm is clamped at the table edge and reaches toward a lower surface, so this may be real, but it may also mean shoulder_lift carries an offset that has not been determined. The rest pose is the only configuration verified against ground truth. Recalibrating both arms in the URDF's zero pose would remove the guesswork entirely.

The elbow_flex mapping is not fitted by eye. It was derived from the wrist camera: at every frame where the camera looks at the ceiling the arm is folded with the gripper up, and where it looks at the desk the arm is extended over it. Checked at 12 frames spanning episode 0, the mapping agrees with the camera at all of them, and it puts 0 of 35,676 frames outside the URDF's joint limits (the untransformed angles put 13,104 outside).

Do not train on this copy or replay it on hardware - the angles no longer match what the servos report. Use ..._40ep_fixed for that.

The original recording had two joints sitting on the STS3215 encoder's 0/4095 wrap point, because the pose held during calibration was not the middle of their real travel. elbow_flex crossed it, so the reading rolled 4095 -> 0 and the recorded angle jumped a full 360 degrees; wrist_roll stopped 43 counts short of it and could not be commanded past it, leaving a ~16 degree gap between commanded and measured position.

Unwrapping elbow_flex is common to both copies and is exact: a wrap is an artefact of the angle representation, so making each episode continuous restores the trajectory (largest within-episode step drops from 360 to 6 degrees).

wrist_roll is not recoverable in either copy. Where it was commanded past the boundary the arm never went there, so no measurement of the intended pose exists. Do not train a policy on that joint - those episodes need re-recording after recalibrating both arms.

Repaired with so101-unwrap-dataset; see SO101_ENCODER_WRAP.md in the LeRobot fork used here.

Dataset Structure

meta/info.json:

{
    "codebase_version": "v3.0",
    "fps": 30,
    "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.front": {
            "dtype": "video",
            "shape": [
                480,
                640,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "is_depth_map": false,
                "video.height": 480,
                "video.width": 640,
                "video.codec": "av1",
                "video.pix_fmt": "yuv420p",
                "video.fps": 30,
                "video.channels": 3,
                "has_audio": false,
                "video.g": 2,
                "video.crf": 30,
                "video.preset": 12,
                "video.fast_decode": 0,
                "video.video_backend": "pyav",
                "video.extra_options": {}
            }
        },
        "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": 40,
    "total_frames": 35676,
    "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": "so_follower",
    "splits": {
        "train": "0:40"
    }
}

Citation

BibTeX:

[More Information Needed]