openarm_monkey_american_336episodes

Joint-space GR00T dataset built from the banana_sung capture on 2026-08-28.

336 episodes / 62,008 frames @ 20 fps / LeRobot v2.1 / robot_type openarm_rh56f1
task: "pick up the banana and put it in the bowl"
state / action  46 columns: head(2) left_arm(7) left_hand(6) right_arm(7)
                            right_hand(6) left_eef(9) right_eef(9)
video           4 views @ 640x360: chest, external, head, wristcam_right

Destination: /media/ssd5/sungjoo/workspace/dataset/openarm_monkey_american_336episodes (= /host_workspace/dataset/openarm_monkey_american_336episodes in the container).

Built by build_american_joint.py, which applies the same corrections, the same constants and the same output format as build_monkey_joint.py — it imports the repository's own parquet writer and stats fingerprint rather than reimplementing them, so it cannot drift from what DatasetWriter produces.

---

Read this before training: the left hand is grafted, not measured

The six left_hand columns are NOT measurements of this session.

The left hand's EtherCAT link was down for the entire capture — master 1 reported Link: DOWN with zero slaves — so the bridge published clamp endpoints instead of measurements: thumb_1 exactly 0 and thumb_2 exactly the URDF maximum, for all 62,008 frames. That is the signature of no feedback, not of a hand at rest. Six of the 26 columns monkey_reboot_joint_config.py trains would have had zero variance, which is a normalisation hazard rather than merely wasted capacity.

The left hand is not used on this rig, so those columns carry no information either way. They were replaced, by graft_left_hand.py, with real left-hand tracks from monkey_reboot_joint_v21 — the same rig, the same 20 fps, the same joint definitions. Episode i takes donor episode i % 330, linearly resampled onto episode i's length; action[t] = state[t+1] follows the dataset's own convention for this group. It is deterministic, so re-running it reproduces the same bytes.

Every episode is marked in meta/episodes.jsonl under vclab.left_hand, and the full episode-to-donor mapping is in meta/left_hand_graft.json. Nothing downstream can mistake these values for recorded data without ignoring both.

The result is at parity with the dataset that is already known to train:

observation.state left_hand	monkey_reboot_joint_v21 (trained, serving)	this dataset
q99 - q01	[0.00267, 0.0, 0.172, 0.53733, 0.17867, 0.50267]	[0.00267, 0.0, 0.172, 0.53733, 0.17867, 0.504]
std	[0.01617, 8e-05, 0.05561, 0.17844, 0.05715, 0.17071]	[0.01645, 6e-05, 0.06911, 0.22066, 0.07118, 0.21142]

thumb_2 has q99 - q01 == 0 in both — it was stuck at its raw count in the donor capture too, and checkpoint-21271 trained through it. GR00T clamps a zero range to 1e-8 in state_action_processor.py, so this is the known-good condition rather than a new hazard.

Because the left hand is trained on grafted tracks, a policy from this dataset will emit left-hand commands. They are meaningless here, and harmless while the left hand stays disconnected; if it is ever reconnected, drop left_hand from the config rather than letting those commands reach the hardware.

left_arm is genuine — parked, but it drifts 0.1217 rad, so it is not constant. head(2) is constant zeros as always and is not named by the training config.

---

What was corrected

Four defects, the same four the previous rebuild fixed:

1. The arm action was a copy of the arm state. action[t] == state[t] bit-for-bit on both arms in the capture's own export, because the arms have no joint-space command — only a 4x4 flange pose — so voc/episode/frames.py fell back to the measured joints. Every action chunk's first step was identically zero. Replaced with action[t] = state[t+1] on left_arm, right_arm and left_hand.
2. The right hand keeps its logged command. It is genuine and leads its own state by the Inspire RH56F1's response time. Training on the hand's future state instead would make the robot lag its own grasp twice.
3. The start-of-episode servo twitch, trimmed. 313 of 336 episodes trimmed, median 8 frames, 2,347 frames total (3.65%), capped at 12 frames = 0.6 s. Thresholded on the measured right-arm joints, because the commanded flange shows no burst at all. Per-episode counts are in meta/trim_report.json.
4. The prompt. The capture carried "banana sung" — the session name, passed straight through from --prompt. Replaced with the string the earlier dataset and the serving path both use.

---

Why this was built from the export, not the raw session

apps/teleop.py numbers raw episodes from 000 on every launch, so relaunching with the same --name overwrote the raw directories of the first 75 exported episodes: 336 exported episodes over 260 raw directories. Building from raw would have yielded 260 and silently dropped 76.

Reading the export loses nothing that matters at 20 fps, because every input the raw path uses is already in it:

• observation.state is the measured joints already resampled to 20 Hz by the same voc/episode/frames.py the raw path would call;
• the right hand's action is the genuine logged command, which the export kept;
• twitch_cut thresholds measured right-arm joints at 20 Hz, exactly what the export holds.

What raw would additionally offer is 88 Hz state and 30 fps video — resampling headroom, not content. The raw video is 640x360, the same as the export. Verified on ep 6 and ep 20 (raw overwritten) against ep 200 and ep 330 (raw intact): the reconstructed action carries the same magnitude of real motion.

The teleop numbering bug is written up at https://app.notion.com/p/3ca86d9ba65c81288236e48150a2458f. It is not fixed yet — until start_index= is added to the run() call in apps/teleop.py, give every teleop launch a new --name.

---

Verifying it

pixi run -e teleop python verify_monkey_joint.py \
    /media/ssd5/sungjoo/workspace/dataset/openarm_monkey_american_336episodes

It checks the action convention on the three replaced blocks, that the right hand did not collapse onto state, frame/episode index integrity, NaNs, and video frame counts against the parquet on a sample of episodes.

As built it reports ALL CHECKS PASSED:

episodes 336  frames 62008  fps 20.0  splits {'train': '0:336'}
tasks: ['pick up the banana and put it in the bowl']
NaN state / action    : 0 / 0
action==state[t+1] left_arm  : worst 0.000e+00  OK
action==state[t+1] right_arm : worst 0.000e+00  OK
action==state[t+1] left_hand : worst 0.000e+00  OK
right_hand action==state in 0 episodes (want 0)
episode length        : min 58 (2.90 s)  median 154  max 1979
constant state columns: ['openarm_head_pitch', 'openarm_head_yaw']

The two head columns are the only constants left, and the training config does not name them. Run it again after the transfer — that is what it is for.