yam-pick-duster — joint-space (LeRobot v3.0)

50 teleoperated demonstrations of a single I2RT YAM arm picking up a duster, recorded in VR. Joint-space state and action, two camera views.

An end-effector-space version of the same 50 takes is published separately as Dimios45/yam-pick-duster-ee — same episodes, same wall-clock spans, different action space and format.

At a glance

Episodes	50
Frames	28,068 @ 25 Hz (18.7 min)
Episode length	396–785 frames (15.8–31.4 s), median 547
Robot	I2RT YAM, 6-DoF + linear_4310 gripper, right arm only
Cameras	top (fixed overhead), right_wrist — both 640×480 RGB, uncropped
Task	"pick up the duster"

Roughly 88% of frames contain motion (min 74%, max 92%) — there are no idle or dead takes in this set.

Schema

observation.state  float32 (7,)  [right_joint_1..6 (rad), right_gripper]   measured
action             float32 (7,)  same layout                               commanded
observation.images.top          video (480, 640, 3)
observation.images.right_wrist  video (480, 640, 3)

Gripper convention: 0 = open, 1 = closed. This is the inverse of i2rt's native normalisation, which is converted at record time.

action is what the teleoperator commanded on that tick; observation.state is what the arm measured. The command leads the measurement by a few ticks, as expected of a position-controlled arm under load.

How it was recorded

Meta Quest controllers → WebXR → differential IK → joint commands, using vr-teleop-kit. Operator holds a grip button to clutch the arm; the trigger drives the gripper.

The arm is commanded at 200 Hz and the dataset is sampled from that loop at 25 Hz. Rate matters: an earlier version commanded at the dataset rate and the arm was visibly jittery, because it received a new joint target only every 1/fps s and the IK's per-tick velocity cap tightened by the same factor. Each episode begins from the same home pose (start poses agree to 0.30 mm across all 50 takes).

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Dimios45/yam-pick-duster")
item = ds[0]
item["observation.state"]              # (7,) joints + gripper
item["observation.images.top"]         # (3, 480, 640) float32 in [0, 1], RGB

Works directly with LeRobot-native policies (ACT, diffusion policy, pi0, SmolVLA) — nothing extra needed.

Training a B-spline diffusion policy on this

The bspline-policy stack reads robomimic HDF5, not LeRobot, and dispatches on the observation keys. Convert with tools/to_robomimic.py from vr-teleop-kit:

python tools/to_robomimic.py --from lerobot \
    --repo-id Dimios45/yam-pick-duster --root <local-root> \
    --output-path yam_joint.hdf5 \
    --crop top_image=42,28,598,414        # optional, see below

That yields obs/joint_pos (N,7), obs/top_image, obs/wrist_image (84×84 RGB) and actions (N,7) — the stack's single_yam_joint format, which needs no rotation conversion and no IK at deployment. Matching shape_meta:

shape_meta: &shape_meta
  obs:
    top_image:   {shape: [3, 84, 84], type: rgb}
    wrist_image: {shape: [3, 84, 84], type: rgb}
    joint_pos:   {shape: [7]}
  action:
    shape: [7]

Cropping

Frames are stored uncropped on purpose, so the crop can be retuned without re-recording. The overhead camera's useful region is roughly x=42, y=28, w=598, h=414 — this drops a corner artefact and the bench rail — but verify it against your own scene.

The wrist camera sees the room above the table horizon (~y=110 at the home pose). That crop is pose-dependent: the horizon moves as the arm pitches, so a fixed rectangle that is clean at one pose can cut into the table at another. Check across your workspace before committing.

Whatever crop you train with must be applied identically at deployment, or the policy sees an input distribution it never saw in training. The converter stamps the crop into the HDF5 attributes so the choice travels with the data.

Licence

Apache-2.0.