ar0s/centrifuge-generalization-reference

This dataset was created using LeRobot.

What this dataset is

Rendered expert episodes of the KUKA centrifuge task on a widened reset distribution, recorded to show what the distribution does to the expert -- including the episodes the expert cannot finish. A failed attempt is recorded as the episode (TaskExpert stops commanding where the attempt died), so a failure here is data, not a gap.

• reset profile: A -- machine translation ±30 mm, machine yaw ±15°, each vial ±30 mm, sampled uniformly per seed by autolab_sim/tasks/centrifuge_layout.py. reference is the distribution the 12/12 full-cycle result was measured on; the other profiles are GENERALIZATION_PLAN.md step 7's stages.
• physics: PHYSX_CONTACT_LAST=1 PHYSX_STABILIZATION=1 PHYSX_DETERMINISM=1 plus the vial's solver iterations -- the settings the load fix needs. Recordings made without them are the pre-fix physics and are not comparable: the reference dataset recorded that way is 3 successes in 12 episodes.
• grip: the jaws close onto the cap, not through it -- CARRY_GRIP = 85, which puts one pad 0.16-0.32mm inside the 16.6mm cap, symmetrically, and leaves the vial 0.03-0.15 deg off the jaw axis. The grasp friction is the cap's own Looks/Grip material (1.2/1.0), re-bound after the spawn and combined with max. The earlier centrifuge-generalization-broken-* datasets were recorded with CARRY_GRIP = 90, i.e. the jaws 5.3mm inside the cap with one pad 4.72mm in and the other 0.13mm, at a cap-pad friction of 0.85; that is what tilts the vial in the grip, and it is why those and these are not comparable either. Measured per step by probes/check_grasp_penetration.py.
• rendering: front-centre and wrist-left cameras, 30 fps, control dt=1/30, physics_dt=1/60. Rendered episodes are their own rollouts; their step counts do not match the headless measurements in HANDOFF.md.
• seeds are the reset draws, so episode i is the i-th (profile, seed) pair listed below. The same seed reproduces the same starting bench.

Episode index, in dataset order:

episode 0  profile reference  seed 0  OK
episode 1  profile reference  seed 1  OK
episode 2  profile reference  seed 2  OK
episode 3  profile reference  seed 3  OK
episode 4  profile reference  seed 4  OK
episode 5  profile reference  seed 5  OK
episode 6  profile reference  seed 6  OK
episode 7  profile reference  seed 7  OK
episode 8  profile reference  seed 8  OK
episode 9  profile reference  seed 9  OK
episode 10  profile reference  seed 10  OK
episode 11  profile reference  seed 11  OK
RECORD_LAYOUT_DONE

Failures by stage are the expert's own assertions: transfer blocked is the lid close never reaching its pose, disturbed is picking one vial moving the other, unload grasp/extraction stalled is the pads missing the cap, tilt is insertion, return is a vial not coming back to where it started.

Dataset Structure

meta/info.json:

{
    "codebase_version": "v3.0",
    "fps": 30,
    "features": {
        "action": {
            "dtype": "float32",
            "names": [
                "ee_x.pos",
                "ee_y.pos",
                "ee_z.pos",
                "ee_wx.pos",
                "ee_wy.pos",
                "ee_wz.pos",
                "gripper.pos"
            ],
            "shape": [
                7
            ]
        },
        "observation.state": {
            "dtype": "float32",
            "names": [
                "ee_x.pos",
                "ee_y.pos",
                "ee_z.pos",
                "ee_wx.pos",
                "ee_wy.pos",
                "ee_wz.pos",
                "gripper.pos"
            ],
            "shape": [
                7
            ]
        },
        "observation.images.wrist_left": {
            "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": {}
            }
        },
        "observation.images.front_center": {
            "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": 12,
    "total_frames": 49514,
    "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": "isaac_kuka",
    "splits": {
        "train": "0:12"
    }
}

Citation

BibTeX:

[More Information Needed]