mpark's picture
Upload README.md with huggingface_hub
05ff768 verified
|
Raw
History Blame Contribute Delete
6.62 kB
metadata
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - robotics
  - lerobot
  - manipulation
  - isaac-lab
  - data-generation
  - MimicGen
configs:
  - config_name: default
    data_files: data/*/*.parquet

This dataset was created using LeRobot.

Dataset Description

Simulated manipulation demonstrations generated by our MimicGen reimplementation on the RoboLab (Isaac Lab) benchmark. One of 12 cells in a generator x task x difficulty grid.

What this is

Generator MimicGen (our reimplementation, not the authors' code)
Task banana
Initial-pose randomization medium — position 50%, yaw ±25°
Episodes 3039 (successes only)
Generation success rate 0.279

Action convention — read this before training

Actions are absolute joint targets in the DROID frame (arm 7-DoF with the j7 mount offset applied) plus a binary gripper in {0, 1}. They are not deltas. Applying a delta transform on top will silently produce a different controller.

The binary gripper matters: a continuous gripper channel overestimates downstream policy performance in our measurements, so it is binarized at collection time.

How the data was produced

Source demos (pi05 policy rollouts) -> MimicGen (object-centric SE(3) retargeting) -> for PGDG cells, those MimicGen episodes are in turn the source for the control-point sampler. Every episode is rolled out in physics and only successes are stored, so the generation success rate above is also the cost driver: episodes that fail consume the same simulation budget as ones that succeed.

Difficulty levels randomize the object's initial placement: position as a fraction of the task's feasible radius, and yaw as a half-angle.

Caveats

  • Only successful episodes are included. Failure modes are not represented here.
  • Success is defined by a K-step settle (hold) criterion in the environment, not a single frame.
  • Observations are two cameras (over-shoulder + wrist) at 270x480, 15 fps, plus proprioceptive state.
  • Quantile statistics (q01/q99) are precomputed, and videos are re-encoded to constant frame rate so that torchcodec-based dataloaders work.
  • This is our reimplementation of the published methods, not the original authors' code. Several hyperparameters are not specified in the papers and were chosen by us.

Visualize

Open in the LeRobot dataset visualizer

  • Homepage: [More Information Needed]
  • Paper: [More Information Needed]
  • License: apache-2.0

Dataset Structure

meta/info.json:

{
    "codebase_version": "v3.0",
    "robot_type": "unknown",
    "total_episodes": 3039,
    "total_frames": 692970,
    "total_tasks": 1,
    "chunks_size": 1000,
    "data_files_size_in_mb": 100,
    "video_files_size_in_mb": 200,
    "fps": 15,
    "splits": {
        "train": "0:3039"
    },
    "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",
            "shape": [
                8
            ],
            "names": {
                "motors": [
                    "j0",
                    "j1",
                    "j2",
                    "j3",
                    "j4",
                    "j5",
                    "j6",
                    "gripper"
                ]
            },
            "fps": 15
        },
        "observation.state": {
            "dtype": "float32",
            "shape": [
                8
            ],
            "names": {
                "motors": [
                    "joint_0",
                    "joint_1",
                    "joint_2",
                    "joint_3",
                    "joint_4",
                    "joint_5",
                    "joint_6",
                    "joint_7"
                ]
            },
            "fps": 15
        },
        "observation.velocity": {
            "dtype": "float32",
            "shape": [
                8
            ],
            "names": null,
            "fps": 15
        },
        "episode_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "frame_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "task_index": {
            "dtype": "int64",
            "shape": [
                1
            ],
            "names": null
        },
        "timestamp": {
            "dtype": "float32",
            "shape": [
                1
            ],
            "names": null
        },
        "next.done": {
            "dtype": "bool",
            "shape": [
                1
            ],
            "names": null,
            "fps": 15
        },
        "observation.images.over_shoulder_left_camera": {
            "dtype": "video",
            "shape": [
                270,
                480,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 270,
                "video.width": 480,
                "video.codec": "avc1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 15.0,
                "video.channels": 3,
                "has_audio": false
            }
        },
        "observation.images.wrist_cam": {
            "dtype": "video",
            "shape": [
                270,
                480,
                3
            ],
            "names": [
                "height",
                "width",
                "channels"
            ],
            "info": {
                "video.height": 270,
                "video.width": 480,
                "video.codec": "avc1",
                "video.pix_fmt": "yuv420p",
                "video.is_depth_map": false,
                "video.fps": 15.0,
                "video.channels": 3,
                "has_audio": false
            }
        }
    }
}

Citation

BibTeX:

[More Information Needed]