DATA_DICTIONARY.md

Data Dictionary — Bigenlight/cube_in_cup_raw

Raw teleoperation recordings for the task "put the cube in the cup", captured via a GELLO leader → UR7e follower setup with two RGB cameras, in a single session on 2026-07-20.

This document describes the RAW release: the original per-take HDF5 signal logs plus the original camera MP4s, exactly as recorded. For a ready-to-train version see the LeRobot dataset cube_in_cup_lerobot_v3.

• Scale: 24 takes (take_*) · 6,226 cam1 frames (6,231 cam2) · 207.6 s (~3.5 min) · ~219 MB
• Every take folder take_NN_YYYYMMDD_HHMMSS/ contains exactly three files: cam1.mp4, cam2.mp4, vectors.h5. No hidden or stray files; no sub-directories.
• Verified read-only against all 24 H5 files and all 48 MP4s; no absolute-path leakage and no PII inside the H5.
• Every figure in this document was measured from the files themselves. The machine-readable version is dataset_stats.json.

---

1. Hardware & recording setup

Component	Spec
Robot (follower)	Universal Robots UR7e — 6-DOF collaborative arm. Joint angles/velocities in radians / rad·s⁻¹. This is the arm that executes and the only arm used as policy input at inference.
Teleoperation (leader)	GELLO — low-cost 3D-printed 6-DOF leader arm. Operator moves GELLO; its joint positions are mapped to UR7e joint targets. Recorded gello_* streams are the leader signal — kept for completeness but NOT observable at inference (the robot cannot see the leader). Do not use gello_* as an input feature.
Gripper	Robotiq 2F-85 two-finger parallel gripper. Continuous normalized command (grip_cmd) plus a continuous measured position (grip_pos), and the leader-side trigger gello_grip.
Camera 1	Intel RealSense D435 — RGB only.
Camera 2	Intel RealSense D435if (a D435 variant) — RGB only.
Video format	1280×720 (720p), 30 fps, MPEG-4 (mpeg4), color yuv420p. Verified identical across all 48 videos. No depth / no IR recorded despite RealSense capability — color stream only. The LeRobot copy re-encodes to AV1; RAW keeps MPEG-4.
Viewpoints	Two fixed viewpoints of different kinds: one RealSense wrist-mounted just above the gripper (eye-in-hand close-up), one on a tripod beside the table (third-person scene view). This release does not assert which is cam1 and which is cam2 — that mapping was not verified. The cam1 ↔ cam2 order is meaningful and must be preserved at deploy time.
Fiducial	An ArUco/AprilTag marker is present on the robot base plate, not on the work surface.
Scene	A light wood-grain tabletop holding exactly two objects: a purple/lavender wooden cube (~5 cm) and a sage-green tapered plastic cup. No distractor objects. Success = cube placed in the cup.

---

2. HDF5 file (vectors.h5) — top-level layout

One file per take. Root has 9 groups (schema identical in all 24 files), each a time-series recorded at its own native rate on its own clock. Each group has a t_rel_s dataset = seconds since take start (starts at 0.0). All datasets are 1-D float64, one array per channel (columnar layout — a channel foo is stored as dataset group/foo, NOT as a 2-D table).

Group	Rows (all 24 takes)	Native rate	What it is
cam1_frames	6,226	30.0 Hz	Timestamp + frame index for each Camera-1 video frame
cam2_frames	6,231	30.0 Hz	Timestamp + frame index for each Camera-2 video frame
command	20,371	~98 Hz	Commanded UR7e joint targets (the action), radians
ur_joint_states	19,968	~97 Hz	Measured UR7e joint state: position, velocity, effort
tcp_pose	19,964	~97 Hz	Measured tool-center-point pose (position + quaternion)
wrench	19,965	~97 Hz	6-axis force/torque at the TCP
gripper	7,575	~36.5 Hz	Gripper command, measured position, leader trigger
gello_joint_states	6,240	30.0 Hz	GELLO leader joint pos + vel (teleop only — not for inference)
synchronized	0 (EMPTY)	—	Intended fused/aligned table; empty in all 24 takes — ignore

Measured rates (all 24 takes)

Two different rate statistics are reported because they disagree, and both are true.

Group	mean rate (Hz)	min–max across takes	median Δt (ms)	5th pct Δt (ms)	max gap (ms)	1 / median Δt (Hz)
cam1_frames	29.97	29.93 – 30.02	33.2	28.1	53.8	30.1
cam2_frames	30.00	29.94 – 30.11	33.2	28.8	65.1	30.1
command	97.95	86.87 – 117.72	13.2	2.9	42.2	75.8
ur_joint_states	97.03	86.93 – 99.93	13.4	2.9	43.7	74.6
tcp_pose	97.05	86.97 – 99.94	13.5	2.9	44.1	74.1
wrench	97.05	86.96 – 99.94	13.4	2.9	43.6	74.6
gripper	36.47	34.91 – 37.69	31.9	2.1	62.0	31.4
gello_joint_states	30.01	29.95 – 30.07	33.2	22.8	62.6	30.1

Mean rate = (N−1)/(t_last − t_first) per take, median across takes. Δt percentiles are pooled over all takes.

Why the two rate columns differ: the robot streams are bursty. The inter-sample interval for command / ur_joint_states / tcp_pose / wrench is bimodal — roughly 5 % of intervals are ~2.9 ms (samples arriving back-to-back) while the median is ~13.3 ms. Averaging over the whole take therefore gives ~98 Hz, while the typical spacing corresponds to ~75 Hz. Logging is timestamp-driven, not fixed-period. Always resample using t_rel_s; never assume a fixed dt and never index-align across streams.

Comparison with the sibling dataset. These robot streams run at ~97–98 Hz, versus ~56–60 Hz in banana_in_pot_raw — about 3.2 robot samples per camera frame here against ~2 there. Same rig, faster logging session.

⚠️ The columns attribute quirk (read this)

Every group carries an HDF5 attribute named columns. It is a single scalar JSON string, e.g. '["t_rel_s", "q1", ...]' — NOT a native list/array. Verified: h5py returns Python str in all 24 files. If a naive consumer does list(grp.attrs["columns"]) expecting a list, it iterates the string character-by-character and you get ['[', '"', 't', '_', 'r', ...] — the "garbled char-by-char" failure. Always json.loads(grp.attrs["columns"]). The correct per-group column lists are given verbatim below and do not depend on the attribute.

---

3. Per-group / per-channel schema

All datasets float64, shape (N,) where N = that group's row count for the take. Ranges are measured across all 24 takes (take_23 included).

cam1_frames — Camera-1 frame timeline

Channel	dtype	Unit	Meaning	Range
t_rel_s	float64	s	Time of this frame, since take start	0 → 16.82
frame_idx	float64	index	0-based frame number in cam1.mp4 (float-typed)	0 → 504

cam2_frames — Camera-2 frame timeline

Channel	dtype	Unit	Meaning	Range
t_rel_s	float64	s	Time of this frame, since take start	0 → 16.80
frame_idx	float64	index	0-based frame number in cam2.mp4 (float-typed)	0 → 503

command — commanded UR7e joint targets ➜ the ACTION

Channel	dtype	Unit	Meaning	min	max	mean
t_rel_s	float64	s	Timestamp	—	—	—
cmd1	float64	rad	Target angle, joint 1 (base)	2.6016	3.2395	2.8682
cmd2	float64	rad	Target angle, joint 2 (shoulder)	−1.7333	−1.1296	−1.4453
cmd3	float64	rad	Target angle, joint 3 (elbow)	1.4592	2.3129	1.9785
cmd4	float64	rad	Target angle, joint 4 (wrist 1)	−2.6559	−1.5630	−2.2316
cmd5	float64	rad	Target angle, joint 5 (wrist 2)	−1.7499	−1.3692	−1.5019
cmd6	float64	rad	Target angle, joint 6 (wrist 3)	−3.7518	−3.0308	−3.4761

ur_joint_states — measured UR7e joint state ➜ core of observation.state

Channel	dtype	Unit	Meaning	min	max	mean
t_rel_s	float64	s	Timestamp	—	—	—
q1	float64	rad	Measured angle, joint 1	2.6016	3.2402	2.8606
q2	float64	rad	Measured angle, joint 2	−1.7067	−1.1299	−1.4411
q3	float64	rad	Measured angle, joint 3	1.4553	2.3131	1.9832
q4	float64	rad	Measured angle, joint 4	−2.6559	−1.5613	−2.2401
q5	float64	rad	Measured angle, joint 5	−1.7499	−1.3693	−1.5022
q6	float64	rad	Measured angle, joint 6	−3.7516	−3.0714	−3.4827
qd1	float64	rad·s⁻¹	Measured velocity, joint 1	−0.6425	0.5499	−0.0186
qd2	float64	rad·s⁻¹	Measured velocity, joint 2	−0.6479	0.5250	−0.0019
qd3	float64	rad·s⁻¹	Measured velocity, joint 3	−0.6256	0.4089	0.0402
qd4	float64	rad·s⁻¹	Measured velocity, joint 4	−0.5344	0.6628	−0.0609
qd5	float64	rad·s⁻¹	Measured velocity, joint 5	−0.2336	0.6557	0.0070
qd6	float64	rad·s⁻¹	Measured velocity, joint 6	−0.6309	0.6635	−0.0340
eff1	float64	N·m (motor effort / current proxy)	Effort, joint 1	−3.8302	3.2727	−0.2260
eff2	float64	N·m	Effort, joint 2	−7.0836	3.2122	−2.1933
eff3	float64	N·m	Effort, joint 3	−4.5974	5.8018	−1.4677
eff4	float64	N·m	Effort, joint 4	−0.8045	1.9341	−0.4613
eff5	float64	N·m	Effort, joint 5	−0.9306	0.8294	0.0757
eff6	float64	N·m	Effort, joint 6	−0.5034	0.5093	−0.0513

tcp_pose — measured tool-center-point pose (base frame)

Channel	dtype	Unit	Meaning	min	max	mean
t_rel_s	float64	s	Timestamp	—	—	—
x	float64	m	TCP position x in robot base frame	0.4199	0.5975	0.5090
y	float64	m	TCP position y	−0.1456	0.1888	−0.0019
z	float64	m	TCP position z	0.1785	0.4873	0.2971
qx	float64	unit quaternion	Orientation x	−0.7475	0.7919	0.2863
qy	float64	unit quaternion	Orientation y	−0.7860	0.8321	0.3238
qz	float64	unit quaternion	Orientation z	−0.1070	0.1458	0.0444
qw	float64	unit quaternion	Orientation w	0.0000	0.1904	0.0571

Workspace extent: ~18 cm in x, ~33 cm in y, ~31 cm in z — a compact tabletop envelope. Note qw stays near 0 while qx/qy dominate: the tool points essentially straight down throughout, as expected for a top-down pick-and-place.

wrench — 6-axis force/torque at TCP

Channel	dtype	Unit	Meaning	min	max	mean
t_rel_s	float64	s	Timestamp	—	—	—
fx	float64	N	Force along x	−17.544	63.420	1.849
fy	float64	N	Force along y	−17.387	58.453	−1.510
fz	float64	N	Force along z	−143.191	5.015	−2.095
tx	float64	N·m	Torque about x	−6.882	2.026	0.518
ty	float64	N·m	Torque about y	−3.852	3.509	−0.394
tz	float64	N·m	Torque about z	−1.996	0.706	0.142

The wide fz excursion (down to −143 N) is a transient contact/acceleration spike; the mean is −2.1 N with a standard deviation of 9.3 N, so the stream sits near zero the overwhelming majority of the time.

gripper — gripper signals

Channel	dtype	Unit	Meaning	min	max	mean
t_rel_s	float64	s	Timestamp	—	—	—
gello_grip	float64	normalized 0–1	Leader (GELLO) grip trigger — teleop only	0.0000	1.0000	0.3787
grip_cmd	float64	normalized 0–1 (open→close)	Commanded gripper (part of the action)	0.0000	0.9998	0.3788
grip_pos	float64	normalized	Measured gripper opening (part of observation.state)	0.0118	0.6314	0.2183

grip_pos never reaches 1.0 — 0.0118 is the closed extreme and 0.6314 the widest observed opening (in take_09; typical open is ~0.50). That is the physical range the fingers swept, not clipping. grip_cmd and gello_grip track each other almost exactly (means 0.3788 vs 0.3787).

gello_joint_states — GELLO leader joints (teleop only, NOT for inference)

Channel	dtype	Unit	Meaning	min	max
t_rel_s	float64	s	Timestamp	—	—
q1	float64	rad	Leader joint 1 angle	2.6001	3.2413
q2	float64	rad	Leader joint 2 angle	−1.7489	−1.1277
q3	float64	rad	Leader joint 3 angle	1.4707	2.3175
q4	float64	rad	Leader joint 4 angle	−2.6611	−1.5673
q5	float64	rad	Leader joint 5 angle	−1.7514	−1.3679
q6	float64	rad	Leader joint 6 angle ⚠ +2π offset	2.5307	3.2578
qd1…qd6	float64	rad·s⁻¹	Leader joint velocities 1–6	—	—

⚠️ gello_q6 is wrapped +2π relative to ur_q6. Joints 1–5 track the follower closely — the mean of gello_qi − ur_qi (nearest-timestamp aligned, pooled over all 24 takes) is within ±0.011 rad for i = 1…5. For joint 6 it is +6.2797 rad ≈ +2π, so gello_q6 lives in [2.53, 3.26] while ur_q6 lives in [−3.75, −3.07]. The two describe the same physical wrist angle in different revolutions. Subtract 2π from gello_q6 before comparing it to the follower or you will read it as a huge tracking error. (This is another reason not to feed gello_* to a policy — but it matters if you study the leader/follower relationship.)

synchronized — EMPTY in all 24 takes (do not use)

The group exists and declares 56 channels, but every dataset has shape (0,) in every take. Fusion was done downstream at conversion time, not stored here. Ignore it. The declared header, in full:

t_rel_s, t_wall,
gello_q1, gello_q2, gello_q3, gello_q4, gello_q5, gello_q6,
gello_qd1, gello_qd2, gello_qd3, gello_qd4, gello_qd5, gello_qd6,
gello_grip,
cmd1, cmd2, cmd3, cmd4, cmd5, cmd6,
ur_q1, ur_q2, ur_q3, ur_q4, ur_q5, ur_q6,
ur_qd1, ur_qd2, ur_qd3, ur_qd4, ur_qd5, ur_qd6,
ur_eff1, ur_eff2, ur_eff3, ur_eff4, ur_eff5, ur_eff6,
grip_cmd, grip_pos,
fx, fy, fz, tx, ty, tz,
tcp_x, tcp_y, tcp_z, tcp_qx, tcp_qy, tcp_qz, tcp_qw,
cam1_frame_idx, cam2_frame_idx

---

4. Per-take statistics

Row counts per stream, duration, and the measured mean command rate. Δcam = cam1 rows − cam2 rows. Every take's video frame counts match its cam*_frames row counts exactly, both cameras.

take	dur (s)	cam1	cam2	Δcam	command	ur_joint_states	tcp_pose	wrench	gripper	gello	cmd Hz
take_01_20260720_205207	16.82	505	504	+1	1718	1662	1661	1662	608	505	102.2
take_02_20260720_205410	16.41	492	493	−1	1617	1617	1617	1617	588	493	98.4
take_03_20260720_205457	10.75	323	323	0	1076	1075	1075	1075	391	323	99.9
take_05_20260720_205546	9.11	273	274	−1	899	867	866	866	334	274	98.5
take_07_20260720_205632	8.66	260	261	−1	807	804	803	803	319	260	93.0
take_08_20260720_205653	8.23	247	248	−1	821	820	820	820	303	248	99.5
take_09_20260720_205713	11.36	341	341	0	1107	1107	1106	1106	421	342	97.5
take_10_20260720_205742	6.88	206	207	−1	648	647	647	647	251	208	93.8
take_11_20260720_205805	7.56	227	226	+1	690	690	690	690	274	227	91.4
take_12_20260720_205835	7.36	221	220	+1	669	669	669	669	270	221	91.1
take_13_20260720_205855	8.29	249	248	+1	883	816	816	816	304	249	106.4
take_14_20260720_205919	6.97	209	210	−1	606	607	607	607	250	209	86.9
take_15_20260720_210005	8.45	253	254	−1	840	833	833	833	310	255	99.1
take_16_20260720_210041	8.67	260	260	0	838	831	831	831	314	260	96.8
take_17_20260720_210110	8.45	253	253	0	779	773	773	773	308	254	92.1
take_18_20260720_210129	7.15	214	215	−1	845	717	717	717	271	216	117.7
take_19_20260720_210149	6.82	204	204	0	683	679	679	679	250	204	100.2
take_20_20260720_210211	8.09	242	243	−1	849	789	789	789	295	244	104.6
take_21_20260720_210234	7.71	232	230	+2	727	727	727	727	281	231	94.5
take_22_20260720_210255	7.37	221	221	0	743	723	723	723	270	222	100.7
⚠️ take_23_20260720_210316	1.64	49	50	−1	170	162	162	162	57	50	103.0
take_24_20260720_210319	8.24	247	247	0	776	775	775	775	300	247	94.2
take_25_20260720_210344	8.96	269	269	0	868	866	866	866	324	268	97.0
take_26_20260720_210403	7.63	229	230	−1	712	712	712	712	282	230	93.3
TOTAL	207.58	6226	6231	—	20371	19968	19964	19965	7575	6240	—

Duration: median 8.24 s, min 1.64 s (take_23, the misfire), max 16.82 s (take_01).

---

5. How to load (h5py)

import json, h5py, numpy as np

path = "Put_cube_in_cup/take_03_20260720_205457/vectors.h5"
with h5py.File(path, "r") as f:
    # correct way to read the column list (do NOT list() the raw attr string):
    cols = json.loads(f["ur_joint_states"].attrs["columns"])   # -> ['t_rel_s','q1',...]

    # measured UR7e joint positions (N_ur, 6), radians, on the UR clock (~97 Hz)
    ur_t = f["ur_joint_states"]["t_rel_s"][:]
    ur_q = np.stack([f["ur_joint_states"][f"q{i}"][:] for i in range(1, 7)], axis=1)

    # commanded joint targets = the action (N_cmd, 6), ~98 Hz
    cmd_t = f["command"]["t_rel_s"][:]
    cmd   = np.stack([f["command"][f"cmd{i}"][:] for i in range(1, 7)], axis=1)

    # camera master timeline (30 Hz); frame_idx maps into cam1.mp4
    cam1_t   = f["cam1_frames"]["t_rel_s"][:]
    cam1_idx = f["cam1_frames"]["frame_idx"][:].astype(int)

# streams are at DIFFERENT rates and arrive in bursts — align to the camera grid
# by nearest timestamp, never by index:
def nearest_idx(src_t, query_t):
    j = np.clip(np.searchsorted(src_t, query_t), 1, len(src_t) - 1)
    left, right = src_t[j - 1], src_t[j]
    return np.where(query_t - left <= right - query_t, j - 1, j)

ur_on_cam = ur_q[nearest_idx(ur_t, cam1_t)]   # (N_frames, 6) aligned to video

Read a specific video frame (OpenCV): cv2.VideoCapture("cam1.mp4") then read sequentially; frame k corresponds to cam1_frames/frame_idx[k]. Because cam1 and cam2 differ by ±1 frame in 16 of 24 takes, map between the two cameras by nearest timestamp too.

6. How to convert to LeRobot

Use convert_to_lerobot.py — the same recipe as the banana family. It resamples every stream onto the cam1 timestamp grid at 30 fps via nearest-timestamp lookup and produces:

• observation.state (7) = ur_joint_states q1..q6 + gripper/grip_pos
• action (7) = command cmd1..cmd6 + gripper/grip_cmd
• observation.images.cam1, observation.images.cam2 (720×1280×3 RGB video, re-encoded to AV1)
• gello_* streams are intentionally dropped (not observable at inference).
• take_23 is excluded, giving 23 episodes from 24 raw takes.

lr_env/bin/python convert_to_lerobot.py --data Put_cube_in_cup \
    --out cube_in_cup_lerobot --repo-id Bigenlight/cube_in_cup_lerobot_v3

---

7. Anomalies & data-quality notes

• ⚠️ take_23_20260720_210316 is a misfire. Duration 1.64 s (49 cam1 frames); the UR is essentially motionless (per-joint range < 0.013 rad); the gripper is never actuated (grip_pos flat at 0.0118, grip_cmd flat at 0.0, total range 0.0). It contains no demonstration of the task. It is included in this raw release (raw = as recorded) and excluded from cube_in_cup_lerobot_v3. Drop it in any pipeline built from these files.
• Take-number gaps {4, 6} are by design. Folder names span take_01…take_26; takes 04 and 06 were aborted/discarded during recording, leaving exactly 24 folders. Not missing data.
• Duration outliers (not defects): take_01 (16.82 s) and take_02 (16.41 s) run about 2× the median take length (8.24 s) — early, slower demonstrations. Their signals are clean.
• Camera frame-count mismatch: in 16 of 24 takes cam1 and cam2 differ by ±1 frame, except take_21 at +2 (totals: 6,226 cam1 vs 6,231 cam2). Per-take deltas are in §4. Handle by nearest-timestamp mapping between the two camera clocks — never assume cam1[k] and cam2[k] are simultaneous.
• synchronized/ empty in all 24 takes (see §3).
• columns attribute is a JSON string, not a list — always json.loads it (see §2).
• Bursty sampling on the robot streams (bimodal Δt, see §2) — resample by timestamp.
• No NaN and no Inf anywhere: every channel of every group of all 24 takes is finite.
• Video ↔ HDF5 agreement is exact: all 24 takes, both cameras (48/48 videos), have MP4 frame counts equal to their cam*_frames row counts (verified with ffprobe -count_frames).
• Timing: largest single gap in any robot stream is 44 ms; largest in any camera stream is 65 ms (~2 frame periods). Max-to-median interval ratio is 1.2–3.2× per take.
• Joint safety envelope: all takes stay within it (see §3 for measured per-joint ranges).
• Uniform schema: identical 9 groups, identical channel names, all float64, in all 24 files.
• Cleanliness: 0 stray files, 0 hidden files, no sub-directories in the take folders, and no /home/ absolute-path strings inside any of the 24 H5 files.

No success/failure labels or human quality ratings ship with this dataset, and no trained policy exists for this task yet.

See dataset_stats.json for exact per-take frame counts, durations, row counts, byte sizes, rate statistics, and per-channel min/max/mean/std.