openarm_banana_all_1072episodes

Three banana captures merged into one 20 fps, 46-column GR00T dataset, built 2026-08-28 by build_all_1072.py.

1,072 episodes @ 20 fps / LeRobot v2.1 / robot_type openarm_rh56f1
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
tasks           0  "pick up the banana and put it in the bowl"    (666 episodes)
                1  "pick up the banana and put it on the plate"   (406 episodes)

output episodes	source	as captured	what changed
0 – 329	openarm_banana_floor2_330episodes	20 fps, 46 col	thumb rescale only
330 – 665	openarm_monkey_american_336episodes	20 fps, 46 col	left-hand thumb rescale only
666 – 1071	openarm_banana_400episode	30 fps, 44 col	resampled, realigned, remapped, thumbs fixed

This is a two-task, two-scene dataset — co-training, not one task with more data. monkey put the banana on a plate in a different room; the other two put it in a bowl. Both prompts are kept and task_index separates them.

---

The seven things that had to be reconciled

Each was measured on the data, not assumed. Every episode records what was done to it in meta/episodes.jsonl under vclab, and meta/merge_report.json has the per-episode frame accounting.

1. Frame rate — monkey was 30, the others 20

monkey's state is linearly interpolated onto the 20 Hz grid and its video is re-encoded at 20 fps. The other two are copied byte for byte, so 2,664 of the 4,288 videos are the originals.

2. Column layout — monkey was 44 columns in a different order

monkey's export puts the end-effector block first (eef, hands, arms); the target is the openarm_rh56f1 order (head, left_arm, left_hand, right_arm, right_hand, left_eef, right_eef). head is two zeros in all three — it was never recorded on this rig.

3. Camera names — the role names invert the mounting

Mapped by physical position, not by name:

monkey	→	here	why
scene_left	→	head	the neck-mounted ZED 2i
ego	→	chest	the static torso ZED Mini
scene_right	→	external	the room view
wrist_left	→	wristcam_right	it is on the right arm

The last row is the one that is easy to get wrong, so it was measured: the camera's own image motion correlates 0.241 with right-arm speed and −0.017 with left-arm speed, and monkey's left arm is parked (median peak-to-peak 0.0078 rad per episode).

---

As built

[build] 1072 episodes, 234978 frames, 195.8 min, 4288 videos
[trim]  monkey episodes trimmed: 231 (median 7 frames)
[check] vendored writer round-trips an existing parquet: schema True  values True

source	episodes	frames	share
floor2 330	330	66,213	28 %
american 336	336	62,008	26 %
monkey 406	406	106,757	45 %

3.8 GB on disk. Episodes 0–999 are in chunk-000, 1000–1071 in chunk-001 — chunks_size is 1000, so this is the first of these datasets to need two.

build_all_1072.py vendors _write_parquet and _fingerprint from voc/episode/lerobot_writer.py because the repository is not on VCL2. It checks itself at startup by round-tripping an existing parquet and refuses to run if the schema or the values differ. Bytes cannot match — pyarrow stamps its own version into the footer and the container's version is not the one the source datasets were written with.

4. Thumb scale — three captures, three affines

[openarm.hand.counts] was corrected twice and both fixes landed between captures, so thumb_1 and thumb_2 were in three different units. Everything is now on the measured affine (552 / 529):

source	thumb_1	thumb_2	four fingers
monkey	750 → ×1.3587	750 → zeroed	750, untouched
floor2	750 → ×1.3587	1400 → ×2.6465	750, untouched
american	already 552	already 529	750, untouched

monkey's thumb_2 is not an angle. At 750 counts/rad that joint's whole command band sat at or above the count the actuator was already resting at, so the thumb was never once asked to bend; the recorded 0.529–0.543 is that resting count seen through the wrong affine, and its true value is 0. Carrying the recorded number into a dataset whose other episodes are on the corrected affine would have commanded a bend that never happened.

The four fingers are on 750 in all three and are still unmeasured — they are mutually consistent, and may be wrong together by the same factor.

5. State-to-video alignment — monkey's state led its video by 0.218 s

The monkey/merged manifests have no clock block, so t.npy's zero was never tied to the camera clock and the archive's video was paired by some other path. Measured two ways on the wrist camera, which is bolted to the arm and therefore sees the arm's own motion — no calibration involved:

                                  mean frame difference   phase correlation
openarm_banana_400episode  30 fps   +6.41 fr = +0.214 s   +6.54 fr = +0.218 s
..._american_336episodes   20 fps   -0.12 fr              -0.01 fr
..._floor2_330episodes     20 fps   -0.05 fr              -0.90 fr

Output frame j therefore takes state at j/20 and video at j/20 + 0.218, done on the continuous timeline so no integer rounding enters. Re-measured on the merged result: +0.0 frames.

6. Right-hand action — monkey's export had thrown the command away

floor2 and american keep the hand's genuine logged command, which leads its own state by a measured 0.40 s (the Inspire RH56F1's response time). monkey's export is action[t] = state[t+1] on all four groups — exactly, worst error 0.0000 over 60 episodes checked. Training one column on two conventions teaches the average of both.

monkey's right-hand action is therefore reconstructed as state[t+8] — the state 0.40 s later, which is what a command at t was asking for. This is a reconstruction, not a recording, and every monkey episode says so: vclab.action_source.right_hand = "reconstructed_state_t+8". Rebuild with --monkey-hand next_state to keep the export's own convention instead.

The three arm groups keep the convention all three sources already shared, action[t] = state[t+1], recomputed after every resample and trim.

7. Prompt — they are different tasks, and are kept that way

monkey's target was a plate in another room. Merging the prompts would claim the scenes are the same; keeping two task_index values makes this co-training, which is what a merged set of two scenes actually is.

---

Also carried over from the sources

• american's left_hand is grafted, not measured — real tracks from floor2, episode i taking donor i % 330. See that dataset's README. Those columns are on floor2's old thumb affine, so they take floor2's rescale here.
• american ep203 (now 533) has a 46-frame right-hand feedback dropout: all six columns pinned at open_counts / counts_per_rad while the action runs a full grasp. It is in the 88 Hz raw too, so it is unrecoverable. 46 frames of 198,000; GR00T N1.7 normalises on q01/q99, which 46 frames cannot move.
• The left arm does nothing in any of the three, and — checked, because a merged set is where this would bite — it does not tag the source. The three park it within 0.173 rad of each other (worst joint, floor2 vs american), and that is smaller than monkey's own spread across its episodes (std 0.09–0.15). The one source that differs in task, monkey, is the least separable of the three. left_arm is in the 26 columns the config trains, so this was worth measuring; it is not a shortcut the model can take to the task label.
• The start-of-episode lunge into the teleop-ready pose is trimmed in all three, by the same twitch_cut with the same constants.

Two things that look wrong and are not

thumb_1 sits just past its URDF limit in 59 % of the floor2 episodes. The excess is median 0.0016 rad, p99 0.0052, max 0.0324. One raw ANGLEACT count is 1/552 = 0.0018 rad, so this is single-count feedback jitter around the mechanical stop — and the URDF limit is that stop. floor2's thumb rests open at raw 600, which the corrected affine maps to 2.0942, two ten-thousandths below the limit; half the samples land on the other side of it. monkey shows the same thing 48 times. Nothing to fix.

The only genuinely out-of-range values in the whole dataset are the 46 frames of the american dropout (now episode 533), which are left in deliberately.

36 monkey episodes have their last video frame twice. ffmpeg's fps filter emits an output frame only when an input frame's display interval covers the output instant, so an episode that ends between two 30 fps frames loses its last 20 fps slot and the video comes out one frame short of the parquet. Caught by verify_merged.py, fixed by re-encoding those 144 videos with tpad=stop_mode=clone and cutting to the row count with -frames:v. The duplicate is the episode's final frame, and the episodes are marked vclab.video_tail_cloned = 1. The builder now does this on every monkey episode.

After training starts, fix meta/relative_stats.json again

GR00T's DatasetFactory regenerates it on rank 0 as root inside the container, mode 600 — overwriting whatever ownership it had. Fixing the permissions before launching does not survive; do it after the run is past dataset setup, or a later server-to-server tar skips the file and still exits 0:

docker exec sungjoo_groot_mr chown 1020:1022 \
  /host_workspace/dataset/openarm_banana_all_1072episodes/meta/relative_stats.json
docker exec sungjoo_groot_mr chmod 644 \
  /host_workspace/dataset/openarm_banana_all_1072episodes/meta/relative_stats.json

Verifying it

python verify_merged.py /host_workspace/dataset/openarm_banana_all_1072episodes

It checks the LeRobot invariants, that task_index matches each episode's source, the action convention on the three arm groups, that the right hand did not collapse onto state, that the thumb columns now sit on one scale, and every video's frame count against its parquet.