openarm_banana_400episode

Bimanual OpenArm teleoperation: pick up the banana and put it on the plate. 406 episodes / 165,349 frames at 30 Hz (91.9 minutes), four camera views, LeRobot v2.1.

episodes	406
frames	165,349 @ 30 fps (91.9 min)
task	single, "pick up the banana and put it on the plate"
robot	OpenArm v10 bimanual, Inspire RH56F1 hands
cameras	ego, scene_left, scene_right, wrist_left — 640×360 h264
state / action	44 dims each, identical block layout
splits	train only; no holdout is carved out

Episode lengths audit clean: shortest 217 frames (7.2 s), median 369 (12.3 s), longest 1,371 (45.7 s), and nothing under 5 s.

Layout

meta/info.json  episodes.jsonl  tasks.jsonl  modality.json  stats.json  build_report.json
data/chunk-000/episode_{000000..000405}.parquet
videos/chunk-000/observation.images.{ego,scene_left,scene_right,wrist_left}/episode_*.mp4

observation.state and action are both 44-wide with the same blocks:

block	slice	meaning
left_eef	0:9	left flange pose, xyz + rot6d
right_eef	9:18	right flange pose, xyz + rot6d
left_hand	18:24	6 driven finger joints
right_hand	24:30	6 driven finger joints
left_arm	30:37	7 arm joints
right_arm	37:44	7 arm joints

Both parameterisations ride in one dataset on purpose: a config that names left_eef, right_eef, left_hand, right_hand trains end-effector control, one that names left_arm, right_arm, left_hand, right_hand trains joint control, and neither reads the other's blocks. Carrying both costs 14 float32 per frame.

Action is the next state

action[t] = state[t+1], on the same 30 Hz grid — not a logged controller command.

The dataset merges two capture sessions. One logs a joint-space command; the other does not, because its arm command survives only as a 4×4 end-effector pose and the joint command was produced by the robot's IK and discarded. Giving 250 episodes one action convention and 156 another would ask a model to unlearn the difference, so every episode uses the next measured state. Where the native command does exist it leads its own state by about 0.3 s, so the two conventions are close.

Actions are stored absolute. Frameworks that want relative actions convert at load time; for an SE(3) end-effector action that is composition in the current flange frame, not a coordinate-wise subtraction.

The EEF frame is the flange, not a fingertip

left_eef / right_eef are link7, the wrist flange, in the shared openarm_body_link0 torso frame — computed by forward kinematics from the joints in the same row, not measured separately.

The robot wears Inspire RH56F1 hands rather than the URDF's parallel gripper, and no flange→palm calibration is published with this data, so serving flange poses keeps uncalibrated constants out of the dataset. A robot consuming these actions must run its IK against the flange and apply its own tool transform.

The 9 numbers are xyz followed by the first two rows of the rotation matrix. Rows and columns differ by a transpose, which silently inverts every rotation and is invisible in a training loss, so check this against your framework's own rot6d helper.

Cross-checked against the openarm_wristcam_bundle v10 model: MuJoCo's own link7 body position agrees with the stored left_eef/right_eef to 0.001 mm, and every one of the 165,349 × 28 arm joint values sits inside the URDF limits.

Hand channels

The 6 per-hand numbers are the independently driven Inspire RH56F1 joints, in order:

thumb_1, thumb_2, index_1, middle_1, ring_1, little_1

The hand's other 6 joints are <mimic> and follow with the URDF's multipliers — thumb_3 = 1.2953·thumb_2, thumb_4 = 0.8962·thumb_3, and {index,middle,ring,little}_2 = 1.1545·{...}_1. They are absent here because nothing drives them independently and the hardware cannot measure them.

Provenance

Two teleoperation sessions on the same rig and the same task, pooled:

session	episodes	frames
A	250	91,645
B	156	73,704

Built from raw capture, not resampled from another release. Every stream carries its own clock — the arms and the wrist camera are written on a remote host that runs 10–12 s behind the main PC — so each episode's per-episode clock offset is applied before anything else, the common window every stream covers is taken, and joints are linearly interpolated onto a 30 Hz grid. Video frames are picked nearest-in-time; pixels are never interpolated. One raw episode of 407 was dropped for having no clock metadata.

Per-episode build records are in meta/build_report.json.

Known limits

• The ego camera is rotated 90°. The scene cameras are upright. Nothing in the pipeline corrects this; rotate it yourself or leave that view out.
• No holdout split. All 406 episodes are in train.
• Videos were re-encoded from the capture at crf 20, measured at ~38 dB PSNR against the source frames.
• People are visible. The scene cameras show the operators. Treat accordingly.

License

Not yet decided — other is a placeholder, not a grant. Ask before redistributing.