DATA_DICTIONARY.md

Data Dictionary — Bigenlight/carrot_in_pot_sim_raw

Raw simulated teleoperation recordings for the task "Put carrot in pot", captured with a physical GELLO leader → MuJoCo UR7e follower and two simulated cameras, in a single session on 2026-09-15 (00:09–00:19 local).

This document describes the RAW sim release: the per-take HDF5 signal logs, the camera MP4s, the exact MJCF scene and the complete per-tick simulator state, exactly as recorded. It is the simulated twin of carrot_in_pot_raw and deliberately uses the same recorder and the same file format. For a ready-to-train version see carrot_in_pot_sim_lerobot_v3.

• Scale: 22 takes (take_*), shipped twice — takes/ (live capture, 9,173 cam1 frames, 9,173 cam2) and retimed_30hz/ (videos re-rendered on an exact 30 Hz grid, 10,725 cam1, 10,725 cam2) · 359.13 s (6.0 min) · 1.32 GB total.
• Every take folder contains exactly three files: vectors.h5, cam1.mp4, cam2.mp4. Measured: no hidden or stray files, no sub-directories, in any of the 44 folders. There is no depth.h5 — depth recording is opt-in in sim_collect and was off for this session.
• Verified read-only against all 44 vectors.h5 and all 88 MP4s. The only absolute path in the release is sim_meta.retimed.from in each retimed_30hz/*/vectors.h5 (source-take provenance); takes/ has none. No PII.
• Every figure here was measured from the files by make_carrot_raw_stats.py (depth-free mode). Machine-readable: dataset_stats.json and retimed_30hz/dataset_stats.json.
• ⚠️ Read §3.6 before you align streams. Unlike the real release, nothing here is stamped late and no constant must be subtracted. Applying the real release's −0.900 s / −0.41 s fix to these files would create a misalignment.

---

1. Simulation & recording setup

Component	Spec
Simulator	MuJoCo 3.10.0, timestep 0.002 s (500 Hz), integrator implicitfast, elliptic friction cone, impratio 10, noslip_iterations 0, arm gravity compensation 1.0
Robot (follower)	Simulated UR7e: mujoco_menagerie ur5e.xml structure with the exact UR7e URDF link offsets (shoulder z 0.1625, wrist_1 z 0.3922, wrist_2 y 0.1263, wrist_3 z 0.0997, attachment_site y 0.0996, elbow ±3.14159). Kinematics DH d = [0.1625, 0, 0, 0.1333, 0.0997, 0.0996], a = [0, −0.425, −0.3922, 0, 0, 0]. attachment_site world pose == ur_kin.fk(q) to 0.000 mm over 2,000 random q. Visual geometry is the UR5e enclosure.
Gripper	mujoco_menagerie Robotiq 2F-85 attached at attachment_site; driver joint robotiq_2f85/right_driver_joint, range 0 – 0.871 rad, normalized 0.0 = open, 1.0 = closed
Teleoperation (leader)	The physical GELLO arm over USB (XL330 ×7, 57600 baud), sampled at 30 Hz, calibrated from the real robot's ROS file ur_gello_bringup/config/ur7e_gello.yaml (gello_publisher). Torque is never enabled.
Control mode	EEF delta, the real ROS bridge code path: One-Euro filter bank (dt = 1/250, min_cutoff 1.0, beta 2.0) → EefDeltaController.step at 250 Hz with pos_scale 1.0, r_align_rpy [0,0,0], tool_l = tool_r = 0.174 m, v_max 0.16 m/s, w_max 1.0 rad/s, max_step_rad 0.0025, soft_start_s 0.7, analytic IK. command = absolute UR joint targets from IK on a leader pose delta.
Gripper mode	continuous, deadband 0.02 (not the 0.3/0.7 discrete latch)
Scene	No table geometry — the workspace is the floor: one 5 m textured plane at z = 0, robot base at the world origin, arm facing world +x at home (home_joints = [−3.302, −1.563, 1.607, −1.523, −1.615, −3.118])
Floor texture	LIBERO seamless_wood_planks_floor.png (MIT, © 2023 Lifelong Robot Learning), texrepeat 20×20, friction [1.0, 0.005, 0.0001]
Objects	carrot (procedural MJCF, food, ~18.5 cm, 0.08 kg) at +y, pot (procedural MJCF, container, inner radius 0.09 m, rim z 0.11 m, floor z 0.006 m, 0.55 kg) at −y
Layout randomization	carrot nominal (0.45, +0.18, 0) ±0.06 m, yaw ±35°; pot nominal (0.45, −0.22, 0) ±0.05 m, yaw ±180°; drop height 0.02 m, settle 0.5 s, min gap 0.02 m, base keep-out radius 0.22 m
Camera 1 — scene	Fixed camera at (0.70, 0.00, 0.571) m, lookat (0.45, 0, 0), up +z, colour fovy 42° → rendered intrinsics fx = fy = 937.83, cx 640, cy 360 at 1280×720
Camera 2 — wrist	Attached to wrist_3_link relative to attachment_site: radial 0.05 m along tool +y, axial 0.08 m, pitch 15° toward the fingertips; same fovy and resolution
Video format	1280×720, 30 fps, mpeg4, yuv420p — measured identical across all 88 videos (ffprobe: mpeg4 1280 720 yuv420p 30/1)
Renderer	MuJoCo offscreen, MUJOCO_GL=glfw, software GL (no NVIDIA driver on the recording machine) — this is why the live capture rate is below 30 fps (§3.1, §9)
Recording software	sim_collect at git commit 4bac8657be5a66adca0b5ccfacc11982420080a1, writing through the real gello_recorder.RecordingSession

cam1 shows the pot, the carrot and the arm entering from the right; cam2 shows the gripper fingers at the bottom of every frame — the same convention as the real release. See assets/take01_cam1_first_frame.jpg, assets/take01_cam2_first_frame.jpg and assets/first_frames.jpg.

---

2. Files in a take folder

File	What it is	Total bytes (takes/, 22)	Total bytes (retimed_30hz/, 22)
vectors.h5	all robot/teleop/gripper/camera-timeline signals + the simulator groups (§3)	99.4 MB	99.8 MB
cam1.mp4	scene camera colour video, 1280×720	266.4 MB	271.8 MB
cam2.mp4	wrist camera colour video, 1280×720	293.0 MB	285.0 MB

Total 658,839,801 bytes (takes/) + 656,642,118 bytes (retimed_30hz/) = 1.32 GB, plus assets/, the two dataset_stats.json, retimed_30hz/retime_manifest.json and this file.

Take names run take_01_20260915_000901 … take_23_20260915_001937. take_06 is absent by design (discarded during recording) — 22 folders, no partial take anywhere. The two sets carry the same 22 take names.

---

3. vectors.h5 — top-level layout

One file per take. Root has 15 groups: the 9 real-recorder groups (schema identical to the real release apart from the extra stamp_s column and a filled synchronized), plus 5 sim_* groups and sim_scene. Every group is a time series at its own native rate on its own clock, with its own t_rel_s = seconds since take start. All datasets are 1-D float64, one array per channel (columnar: channel foo is dataset group/foo, NOT a 2-D table). The one exception is sim_scene/xml, a scalar string dataset.

Group	Rows (22 takes)	Native rate	What it is
cam1_frames	9,173	25.5 Hz live / 30.0 Hz retimed	Timestamp + frame index for each cam1.mp4 frame
cam2_frames	9,173	25.5 Hz live / 30.0 Hz retimed	Timestamp + frame index for each cam2.mp4 frame
command	44,660	125.5 Hz	Commanded UR joint targets (the action), radians
ur_joint_states	44,660	125.5 Hz	Simulated joint state: position, velocity, effort
tcp_pose	44,660	125.5 Hz	TCP pose in base frame (position + quaternion)
wrench	44,660	125.5 Hz	Flange 6-axis force/torque sensor, tared at take start
gripper	22,336	62.8 Hz	Gripper command, measured position, leader trigger
gello_joint_states	10,797	30.2 Hz	GELLO leader joint pos + finite-difference vel
synchronized	33,848	100 Hz	56-channel fused table — ⚠️ FILLED here, empty in the real release (§3.5)
sim_object_poses	10,650	30 Hz	Ground-truth carrot + pot pose (§3.7)
sim_control	44,660	125 Hz	Teleop state machine, IK conditioning, per-tick task success (§3.7)
sim_leader_filtered	44,660	125 Hz	One-Euro filter output the controller consumed (§3.7)
sim_mj_state	44,660	125 Hz	Full generalized state qpos/qvel/ctrl (§3.7)
sim_frame_capture	18,346	~51 Hz	Which physics tick each rendered frame came from (§3.7)
sim_scene	22	once / take	The exact compiled MJCF + asset manifest + layout + config (§3.8)

3.1 Measured rates

Group	mean rate (Hz)	min–max across takes	median Δt (ms)	5th pct Δt (ms)	max gap (ms)	1 / median Δt (Hz)
cam1_frames	25.53	23.71 – 27.31	37.3	25.4	117.4	26.8
cam2_frames	25.47	23.42 – 27.27	37.3	25.5	112.9	26.8
command	125.50	120.30 – 125.86	7.8	0.8	79.3	128.2
ur_joint_states	125.50	120.30 – 125.86	7.8	0.8	79.4	128.2
tcp_pose	125.50	120.30 – 125.86	7.8	0.8	79.3	128.2
wrench	125.50	120.30 – 125.86	7.8	0.8	79.3	128.2
gripper	62.79	60.19 – 63.02	15.9	5.2	96.1	62.9
gello_joint_states	30.18	30.13 – 30.29	32.9	22.9	104.9	30.4
cam1_frames (retimed)	30.00	30.00 – 30.00	33.3	33.3	33.4	30.0
cam2_frames (retimed)	30.00	30.00 – 30.00	33.3	33.3	33.4	30.0

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

The robot streams run at 125 Hz, not the real rig's ~67 Hz. The sim writes ur_joint_states / command / tcp_pose / wrench on every 2nd 2 ms physics step, so there are ~4.2 robot samples per retimed camera frame against ~2.2 on the real rig. The gripper table is written at half that (62.8 Hz) and the leader table at its 30 Hz USB sampling rate.

The live camera rate is the one real defect in this release. cam1_frames / cam2_frames in takes/ average 25.53 / 25.47 Hz (range 23.4 – 27.3) while the MP4 is stamped 30 fps, so the live videos play ~1.18× fast. The timestamps are correct; only the container rate is wrong. retimed_30hz/ fixes it by re-rendering: 30.00 Hz in every take, max Δt 33.4 ms. See §9.

3.2 ⚠️ The columns attribute quirk (inherited from the real recorder)

Every group carries an HDF5 attribute named columns which is a single scalar JSON string, not a native list — measured Python str in all 44 files (dataset_stats.json → aggregate.columns_attr_python_types == ["str"]). list(grp.attrs["columns"]) iterates it character-by-character and yields garbage. Always json.loads(...). The correct column lists are given verbatim below and do not depend on the attribute.

3.3 Per-group / per-channel schema — the nine real groups

All datasets float64, shape (N,). Ranges are measured pooled across all 22 takes of takes/ (the retimed set differs only in cam*_frames).

cam1_frames / cam2_frames — camera frame timelines

Channel	dtype	Unit	Meaning	Range
t_rel_s	float64	s	Time of this frame, since take start	0 → 27.82
frame_idx	float64	index	0-based frame number in the MP4 (float-typed)	0 → 698 live, 0 → 831 retimed
stamp_s	float64	s (unix epoch)	Wall-clock time the frame was captured off the renderer	~1.789e9

⚠️ In retimed_30hz/ these two tables are rewritten onto the exact 30 Hz grid and are the only tables that change. Everything else is byte-for-byte the live recording.

command — commanded UR joint targets ➜ the ACTION

Channel	dtype	Unit	Meaning	Range	Mean
cmd1	float64	rad	Absolute target for joint 1 from EEF-delta IK	-4.0538 → -2.8535	-3.3385
cmd2	float64	rad	Absolute target for joint 2 from EEF-delta IK	-2.0504 → -1.0596	-1.5070
cmd3	float64	rad	Absolute target for joint 3 from EEF-delta IK	1.4157 → 2.6139	1.9385
cmd4	float64	rad	Absolute target for joint 4 from EEF-delta IK	-2.9262 → -1.3392	-1.9637
cmd5	float64	rad	Absolute target for joint 5 from EEF-delta IK	-2.0771 → -1.3681	-1.6389
cmd6	float64	rad	Absolute target for joint 6 from EEF-delta IK	-4.0435 → -2.6561	-3.1377
t_rel_s	float64	s	Time of this sample	0 → 27.82	—

ur_joint_states — simulated joint state ➜ core of observation.state

Channel	dtype	Unit	Meaning	Range	Mean
q1	float64	rad	Measured joint 1 position (d.qpos)	-4.0390 → -2.8625	-3.3338
q2	float64	rad	Measured joint 2 position (d.qpos)	-2.0367 → -1.0833	-1.5041
q3	float64	rad	Measured joint 3 position (d.qpos)	1.4209 → 2.6115	1.9395
q4	float64	rad	Measured joint 4 position (d.qpos)	-2.8825 → -1.3502	-1.9604
q5	float64	rad	Measured joint 5 position (d.qpos)	-2.0592 → -1.3709	-1.6391
q6	float64	rad	Measured joint 6 position (d.qpos)	-4.0384 → -2.6656	-3.1348
qd1..qd6	float64	rad/s	Joint velocities (d.qvel[:6])	-0.5612 → 0.5569 (pooled)	≈0
eff1..eff6	float64	N·m	Actuator force (d.actuator_force[:6])	-64.0041 → 107.1025 (pooled)	—
t_rel_s	float64	s	Time of this sample	0 → 27.82	—
stamp_s	float64	s (unix epoch)	Wall-clock time the physics thread published the tick	~1.789e9	—

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

Channel	dtype	Unit	Meaning	Range	Mean
x	float64	m	TCP x — fk(q) ⊕ 0.174 m along flange +Z	0.2940 → 0.5914	0.4445
y	float64	m	TCP y — fk(q) ⊕ 0.174 m along flange +Z	-0.3670 → 0.2682	0.0268
z	float64	m	TCP z — fk(q) ⊕ 0.174 m along flange +Z	-0.0341 → 0.3300	0.1476
qw	float64	—	quaternion w — fk(q) ⊕ 0.174 m along flange +Z	-0.1283 → 0.1700	0.0029
qx	float64	—	quaternion x — fk(q) ⊕ 0.174 m along flange +Z	0.4296 → 0.9395	0.7617
qy	float64	—	quaternion y — fk(q) ⊕ 0.174 m along flange +Z	0.3219 → 0.9013	0.6241
qz	float64	—	quaternion z — fk(q) ⊕ 0.174 m along flange +Z	-0.3042 → 0.0954	-0.0444
t_rel_s, stamp_s	float64	s	as above	—	—

wrench — 6-axis force/torque at the flange

Channel	dtype	Unit	Meaning	Range	Mean
fx	float64	N	Simulated <force>/<torque> sensor at ft_site, tared at take start	-85.8618 → 91.0505	0.0651
fy	float64	N	Simulated <force>/<torque> sensor at ft_site, tared at take start	-10.6378 → 171.2164	1.1457
fz	float64	N	Simulated <force>/<torque> sensor at ft_site, tared at take start	-9.0993 → 279.1303	0.9739
tx	float64	N·m	Simulated <force>/<torque> sensor at ft_site, tared at take start	-25.8313 → 2.2475	-0.1045
ty	float64	N·m	Simulated <force>/<torque> sensor at ft_site, tared at take start	-16.0800 → 23.2892	0.0140
tz	float64	N·m	Simulated <force>/<torque> sensor at ft_site, tared at take start	-4.1689 → 4.9055	0.0070
t_rel_s, stamp_s	float64	s	as above	—	—

⚠️ Unvalidated against the real arm. The frame and sign convention of this sensor has never been measured against the real UR7e's wrench. Magnitudes are much larger than the real rig's because a 2 ms-timestep contact transient is stiff. Treat as a sim signal.

gripper — gripper signals (0.0 = open, 1.0 = closed, all three)

Channel	dtype	Unit	Meaning	Range	Mean
grip_pos	float64	normalized	Measured normalized opening of the 2F-85 driver joint	0.0000 → 0.9150	0.3367
grip_cmd	float64	normalized	Commanded normalized opening ➜ the gripper action	0.0000 → 0.9998	0.4046
gello_grip	float64	normalized	Raw GELLO trigger value	0.0000 → 1.0000	0.4048
t_rel_s	float64	s	Time of this sample	0 → 27.82	—

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

Channel	dtype	Unit	Meaning	Range
q1	float64	rad	Leader joint 1, unwrapped	-1.0324 → 0.2777
q2	float64	rad	Leader joint 2, unwrapped	-1.9269 → -0.8101
q3	float64	rad	Leader joint 3, unwrapped	1.0642 → 2.1917
q4	float64	rad	Leader joint 4, unwrapped	-3.0860 → -1.6394
q5	float64	rad	Leader joint 5, unwrapped	-1.9248 → -0.8586
q6	float64	rad	Leader joint 6, unwrapped	-1.7614 → 0.9108
qd1..qd6	float64	rad/s	Finite-difference leader velocity	-2.2226 → 1.9564
t_rel_s	float64	s	Time of this sample	0 → 27.82
stamp_s	float64	s (MONOTONIC, not unix)	Leader thread's own clock — ⚠️ a different timebase from every other stamp_s	1715.1 → 2362.5

3.4 The leader stream does not mirror the follower

Same mechanism as the real EEF-mode release: command comes from IK on a leader pose delta, so gello_q* is one IK branch of a virtual leader chain and ur_q* is another solution of a different chain. They are related only through the end-effector pose delta, never joint-by-joint. gello_* is leader-frame telemetry with its own zeros and sign conventions; it cannot be used to reconstruct the action or the follower state, and a deployed policy cannot see it anyway. Use command for the action and ur_joint_states for the state. Per-joint correlations are in dataset_stats.json → leader_vs_follower.

3.5 synchronized — FILLED here (the real release's is empty)

A 56-channel fused table on a 100 Hz grid, 33,848 rows over the 22 takes, starting only once both cameras have delivered their first frame. Channels: t_rel_s, t_wall, cmd1..6, ur_q1..6, ur_qd1..6, ur_eff1..6, tcp_{x,y,z,qw,qx,qy,qz}, fx,fy,fz,tx,ty,tz, gello_q1..6, gello_qd1..6, gello_grip, grip_cmd, grip_pos, cam1_frame_idx, cam2_frame_idx.

🪤 In retimed_30hz/, cam1_frame_idx / cam2_frame_idx still point at the ORIGINAL live frame numbers. Only cam*_frames was rewritten. Use synchronized for camera indexing with takes/ only; with the retimed videos, index through cam*_frames.

3.6 ⚠️ Timestamps — nothing is stamped late, do not shift anything

The real release carries a recorder artefact: three groups stamped 0.900 s / ≈0.41 s late by rclpy spin-thread starvation. That mechanism does not exist here. The sim recorder stamps every robot row with the physics tick that produced it, and command, ur_joint_states, tcp_pose and wrench all come out of the same tick — their t_rel_s agree to within 0.3 ms. Speed cross-correlation between them measures 0.000 s in all 22 takes.

dataset_stats.json → timestamp_lag nevertheless reports a ur_joint_states_lag_s of 0.190 – 0.205 s (median 0.198 s), with the joint residual falling from 0.0187 rad uncorrected to 0.0037 rad at that shift. That is not a clock error — it is the simulated arm's mechanical tracking lag behind its commanded joint target, bounded by the 250 Hz controller's max_step_rad = 0.0025 slew limit and the actuator gains. It is real behaviour of the simulated plant, not an artefact to remove. The JSON says so itself: timestamp_lag.applies_to_this_release = false and timestamp_lag.simulation_note; the note / method prose in that block is inherited verbatim from the real release's script and describes the real rig.

3.7 Simulation-only groups (sim_*)

sim_object_poses — ground truth, 30 Hz

Channel	dtype	Unit	Meaning
carrot_x, carrot_y, carrot_z	float64	m	Carrot body origin in world frame
carrot_qx, carrot_qy, carrot_qz, carrot_qw	float64	—	Carrot orientation
pot_x … pot_qw	float64	m / —	Same for the pot
t_rel_s	float64	s	Time of this sample

Row 0 of this table is the authoritative initial layout of the take — sim_scene's layout attribute is stale (§9). Measured across the 22 takes: carrot x 0.396 – 0.492 m, y +0.124 – +0.242 m; pot x 0.405 – 0.489 m, y −0.191 – −0.258 m; both yaws vary freely.

sim_control — teleop state machine, 125 Hz

Channel	dtype	Meaning
engaged	float64	1.0 while the operator holds the deadman and the controller drives the arm
eef_state_code	float64	0 DISENGAGED · 1 ENGAGING · 2 ENGAGED · 3 HOLD · 4 SOFT_START · 5 REJECTED · 6 FAULT · 7 PAUSED (also in sim_meta.eef_state_codes)
pos_scale	float64	Effective leader→follower position scale (1.0 throughout)
sigma_min	float64	Smallest singular value of the Jacobian — IK conditioning
gamma, ls_scale	float64	Damping / line-search scale inside the IK step
task_success	float64	1.0 on ticks where the geometric success test passes
sim_t, tick	float64	Simulator time (s) and physics tick index
t_rel_s	float64	Time of this sample

Measured over the release: engaged is 1.0 for 71.8 – 100 % of ticks per take (median 100 %); the state histogram is 43,022 ENGAGED · 1,482 HOLD · 156 DISENGAGED ticks, with no REJECTED or FAULT tick anywhere. Every take reaches task_success — first at 8.68 s, median 12.26 s, last at 25.28 s.

The success test (sim_collect/task.py): the food body's origin is inside the container's opening cylinder (above pot_floor, below pot_opening, within inner_radius 0.09 m) and its vertical speed is < 0.05 m/s and grip_cmd < 0.3. It is a geometric approximation for the operator's on-screen badge, not a curated training label.

sim_leader_filtered — 125 Hz

qf1..qf6 (rad) — the One-Euro filter output the EefDeltaController actually consumed, plus t_rel_s. With gello_joint_states this lets you replay the controller offline and reproduce command.

sim_mj_state — the full generalized state, 125 Hz

qpos0..qpos27, qvel0..qvel25, ctrl0..ctrl6, sim_t, tick, t_rel_s. Group attrs: nq = 28, nv = 26, nu = 7, plus a ctrl_note. This is what makes the release replayable — see §5.

sim_frame_capture — render provenance, ~51 Hz (both cameras interleaved)

cam (1 or 2), frame_idx, seq, sim_t, tick, t_capture_rel_s, t_rel_s. Tells you which physics tick each live rendered frame came from. Unchanged in retimed_30hz/, where it therefore still describes the live frames.

3.8 /sim_scene — the exact model

Item	Type	Content
xml	scalar string dataset (~50 kB)	The exact MJCF the simulator compiled for this take
xml_sha256	attr	sha256 of that XML
assets_manifest	attr (JSON)	{asset_filename: {sha256, bytes}} for all 29 referenced meshes/textures. The bytes are not copied into the take — they live in the sim_collect repo at git_commit
layout	attr (JSON)	⚠️ stale — see §9. Per-object pos/yaw/fallback plus _seed, _attempt
layout_seed	attr	⚠️ -1 in all 22 takes (§9)
config	attr (JSON)	The complete scene config: robot, physics, floor, cameras, render, objects, layout, task, control, leader, gripper
config_path	attr	sim_collect/configs/carrot_in_pot_sim.yaml
git_commit, mujoco_version, timestep	attrs	4bac8657…, 3.10.0, 0.002

3.9 File attribute sim_meta

One JSON string on the root of vectors.h5 (the real recorder writes no file attrs at all; adding them breaks no consumer). Keys: sim_collect_version, git_commit, mujoco_version, robot, control_mode, gripper_mode, pos_scale, capture_fps, scene_meta, scene_sha, config, scene_config, objects, chosen_food, container, layout_seed (⚠️ §9), cameras (pose at start + colour/depth render intrinsics per camera), eef_state_codes, record_depth (false), depth_camera_info / depth_extrinsics_depth_to_color / depth_source (the real D435 sidecar values, kept so sim and real takes share one schema even when depth is off), simulated (true), take_name, take_index, started_at, stopped_at, duration_s, camera_fps, sample_rate_hz, recorder_counts, message_counts, achieved_fps, achieved_fps_take, capture, problems, events, object_names, task_success_at_stop. In retimed_30hz/ there is one extra key, retimed: {from, fps, frames, grid_start_t_rel_s, max_state_lookup_dt_s, note, original_problems}.

---

4. Complete pooled value ranges

Every value channel of takes/, pooled over all 22 takes. The stamp_s and frame_idx clocks are omitted (they are indices, not values); all of them are in dataset_stats.json → value_ranges.

Channel	min	max	mean	std
command/cmd1	-4.0538	-2.8535	-3.3385	0.3085
command/cmd2	-2.0504	-1.0596	-1.5070	0.1769
command/cmd3	1.4157	2.6139	1.9385	0.1903
command/cmd4	-2.9262	-1.3392	-1.9637	0.3142
command/cmd5	-2.0771	-1.3681	-1.6389	0.1111
command/cmd6	-4.0435	-2.6561	-3.1377	0.2375
gello_joint_states/q1	-1.0324	0.2777	-0.3984	0.3341
gello_joint_states/q2	-1.9269	-0.8101	-1.3865	0.1902
gello_joint_states/q3	1.0642	2.1917	1.7367	0.2008
gello_joint_states/q4	-3.0860	-1.6394	-2.2805	0.3149
gello_joint_states/q5	-1.9248	-0.8586	-1.4892	0.1773
gello_joint_states/q6	-1.7614	0.9108	-0.3210	0.5730
gello_joint_states/qd1	-1.1027	1.0749	0.0203	0.1746
gello_joint_states/qd2	-1.1159	0.8295	0.0035	0.1655
gello_joint_states/qd3	-0.8712	0.9026	-0.0033	0.1490
gello_joint_states/qd4	-0.8537	1.4665	0.0043	0.2265
gello_joint_states/qd5	-1.1402	1.2635	-0.0134	0.1294
gello_joint_states/qd6	-2.2226	1.9564	0.0227	0.3238
gripper/gello_grip	0.0000	1.0000	0.4048	0.4852
gripper/grip_cmd	0.0000	0.9998	0.4046	0.4839
gripper/grip_pos	0.0000	0.9150	0.3367	0.4031
tcp_pose/qw	-0.1283	0.1700	0.0029	0.0608
tcp_pose/qx	0.4296	0.9395	0.7617	0.0925
tcp_pose/qy	0.3219	0.9013	0.6241	0.1137
tcp_pose/qz	-0.3042	0.0954	-0.0444	0.0558
tcp_pose/x	0.2940	0.5914	0.4445	0.0466
tcp_pose/y	-0.3670	0.2682	0.0268	0.1592
tcp_pose/z	-0.0341	0.3300	0.1476	0.0970
ur_joint_states/eff1	-64.0041	30.0950	0.0293	2.1257
ur_joint_states/eff2	-11.0136	107.1025	-4.4148	4.7394
ur_joint_states/eff3	-13.2352	31.3021	-4.5149	1.5355
ur_joint_states/eff4	-18.0977	10.5806	-0.9808	0.7682
ur_joint_states/eff5	-27.0023	28.0000	-0.0896	1.1235
ur_joint_states/eff6	-4.6147	4.9129	0.0185	0.2001
ur_joint_states/q1	-4.0390	-2.8625	-3.3338	0.3055
ur_joint_states/q2	-2.0367	-1.0833	-1.5041	0.1743
ur_joint_states/q3	1.4209	2.6115	1.9395	0.1899
ur_joint_states/q4	-2.8825	-1.3502	-1.9604	0.3121
ur_joint_states/q5	-2.0592	-1.3709	-1.6391	0.1087
ur_joint_states/q6	-4.0384	-2.6656	-3.1348	0.2339
ur_joint_states/qd1	-0.4522	0.3217	-0.0232	0.1258
ur_joint_states/qd2	-0.4105	0.4322	-0.0035	0.1260
ur_joint_states/qd3	-0.4947	0.4158	0.0063	0.1191
ur_joint_states/qd4	-0.5591	0.5569	-0.0065	0.2213
ur_joint_states/qd5	-0.3949	0.4152	0.0017	0.0895
ur_joint_states/qd6	-0.5612	0.5366	-0.0146	0.1229
wrench/fx	-85.8618	91.0505	0.0651	4.4609
wrench/fy	-10.6378	171.2164	1.1457	6.7999
wrench/fz	-9.0993	279.1303	0.9739	11.6699
wrench/tx	-25.8313	2.2475	-0.1045	0.9711
wrench/ty	-16.0800	23.2892	0.0140	0.8081
wrench/tz	-4.1689	4.9055	0.0070	0.1627

---

5. Reconstruction — rebuilding and replaying a take

/sim_scene + sim_mj_state reproduce every recorded instant kinematically exactly. The replay tool sim_collect/tools/replay_take.py:

1. reads config + layout and rebuilds the scene to obtain the asset files;
2. checks every asset against its recorded sha256 in assets_manifest;
3. compiles the stored xml (not a regenerated one) against those assets — reporting rebuilt xml matches: True;
4. for each sim_mj_state row writes qpos/qvel and calls mj_forward — no physics is re-simulated, so there is no divergence;
5. optionally renders any camera at any stride.

# in a checkout of the recording repo at sim_meta.git_commit (4bac865)
.venv/bin/python -m sim_collect.tools.replay_take <take_dir> --check
#  -> model: nq 28 nv 26 nu 7 | rows 3002 | rebuilt xml matches: True | assets 29
#     object pose reconstruction error (replayed vs recorded sim_object_poses): max 0.00 mm

MUJOCO_GL=glfw DISPLAY=:0 .venv/bin/python -m sim_collect.tools.replay_take <take_dir> --viewer
MUJOCO_GL=glfw DISPLAY=:0 .venv/bin/python -m sim_collect.tools.replay_take <take_dir> \
    --render cam1 cam2 --out /tmp/frames --every 15

Measured object-pose reconstruction error: 0.00 mm, including on takes whose stored layout attribute is stale — because step 4 overwrites qpos regardless. retimed_30hz/ was produced exactly this way.

---

6. Camera timelines and the colour videos

cam*.mp4 are 1280×720 mpeg4 yuv420p stamped 30 fps in both sets. Map frames to signals through cam*_frames:

	takes/ (live)	retimed_30hz/
cam1 frames	9,173	10,725
cam2 frames	9,173	10,725
mean frame rate	25.53 Hz (23.71 – 27.31)	30.00 Hz every take
median Δt / max Δt	37.3 ms / 117.4 ms	33.3 ms / 33.4 ms
takes with cam1 ≠ cam2 count	20 of 22 (−9 … +5)	0
sim_meta.problems	2 entries per take	[]

retime_manifest.json records, per take, the output frame count and max_state_lookup_dt_s — the largest distance between a 30 Hz grid instant and the nearest recorded 125 Hz state row. Worst over all 22 takes: 37.6 ms; take_01 is 19.6 ms.

Video frame counts match cam*_frames row counts exactly in all 88 videos, verified with ffprobe -count_frames.

---

7. Per-take statistics

Take	Duration (s)	live cam1/cam2	retimed cam	live cam rate (Hz)	robot rate (Hz)	gripper closures	max grip_pos
take_01_20260915_000901	15.72	395 / 397	469	25.33 / 25.38	125.7	1	0.7897
take_02_20260915_000933	15.66	403 / 402	468	25.86 / 25.78	124.0	1	0.9132
take_03_20260915_001003	15.96	400 / 400	477	25.18 / 25.14	123.0	1	0.9119
take_04_20260915_001042	16.96	446 / 450	507	26.43 / 26.62	125.5	2	0.9140
take_05_20260915_001112	11.66	314 / 312	348	27.09 / 26.90	125.5	1	0.5652
take_07_20260915_001145	12.68	308 / 307	379	24.38 / 24.37	125.2	1	0.9141
take_08_20260915_001209	24.01	612 / 611	717	25.61 / 25.54	125.5	2	0.9121
take_09_20260915_001242	27.82	699 / 700	832	25.19 / 25.20	124.2	2	0.9131
take_10_20260915_001322	12.28	299 / 297	367	24.44 / 24.34	120.3	1	0.9140
take_11_20260915_001346	12.16	314 / 318	363	26.03 / 26.30	124.8	1	0.9150
take_12_20260915_001405	20.57	537 / 540	615	26.19 / 26.39	124.9	1	0.9144
take_13_20260915_001438	13.01	349 / 353	388	26.96 / 27.27	125.7	1	0.9113
take_14_20260915_001502	21.13	552 / 547	631	26.25 / 26.05	125.6	2	0.9117
take_15_20260915_001534	12.84	342 / 337	383	26.84 / 26.34	125.7	1	0.9119
take_16_20260915_001558	14.02	380 / 377	418	27.31 / 27.01	125.6	1	0.9138
take_17_20260915_001620	22.85	580 / 589	683	25.47 / 25.86	124.4	2	0.9135
take_18_20260915_001658	11.30	283 / 284	337	25.17 / 25.25	122.8	1	0.9106
take_19_20260915_001725	16.37	417 / 413	489	25.59 / 25.36	125.5	1	0.5671
take_20_20260915_001805	19.63	496 / 497	586	25.38 / 25.40	125.6	2	0.9130
take_21_20260915_001833	19.35	489 / 489	578	25.38 / 25.37	125.5	2	0.9132
take_22_20260915_001910	11.86	291 / 290	353	24.73 / 24.61	125.9	1	0.9123
take_23_20260915_001937	11.29	267 / 263	337	23.71 / 23.42	125.6	1	0.9127

Duration: median 15.69 s, min 11.29 s (take_23_20260915_001937), max 27.82 s (take_09_20260915_001242).

---

8. How to load (h5py)

import h5py, json, numpy as np, cv2

take = "retimed_30hz/take_03_20260915_001003"   # train on the retimed set
with h5py.File(f"{take}/vectors.h5", "r") as f:
    ur_q   = np.stack([f["ur_joint_states"][f"q{k+1}"][:] for k in range(6)], 1)  # (N125,6) rad
    cmd    = np.stack([f["command"][f"cmd{k+1}"][:]       for k in range(6)], 1)  # (N125,6) rad
    grip   = f["gripper"]["grip_pos"][:]                                          # (N63,)
    cam1_t = f["cam1_frames"]["t_rel_s"][:]                                       # 30 Hz master grid
    ur_t   = f["ur_joint_states"]["t_rel_s"][:]      # NO shift - see 3.6

    # simulator-only ground truth
    carrot = np.stack([f["sim_object_poses"][f"carrot_{k}"][:] for k in "xyz"], 1)
    succ   = f["sim_control"]["task_success"][:]
    qpos   = np.stack([f["sim_mj_state"][f"qpos{i}"][:]
                       for i in range(f["sim_mj_state"].attrs["nq"])], 1)
    mjcf   = f["sim_scene"]["xml"][()]               # the exact model, as bytes
    meta   = json.loads(f.attrs["sim_meta"])

# streams are at DIFFERENT rates - align to the camera grid by nearest timestamp,
# never by index:
ur_on_cam = np.stack([np.interp(cam1_t, ur_t, ur_q[:, k]) for k in range(6)], 1)

cap = cv2.VideoCapture(f"{take}/cam1.mp4")        # 1280x720 colour @ 30 fps

---

9. Anomalies & data-quality notes

1. 🐛 layout_seed / layout metadata is stale in all 22 takes. sim_meta.layout_seed says 0, /sim_scene attr layout_seed says -1, and /sim_scene attr layout carries one single layout (carrot (0.480, +0.188), pot (0.488, −0.243)) in every take — while the scene really was re-randomized between takes. A recorder bug, fixed after this session. Read sim_object_poses row 0 (or sim_mj_state qpos) for the true initial layout. take_02/take_03 share a layout (no reset between them), as do take_22/take_23. Reconstruction is unaffected (§5): replay overwrites qpos, so the error stays 0.00 mm.
2. ⚠️ synchronized/cam*_frame_idx in retimed_30hz/ refers to the ORIGINAL live frames (§3.5).
3. ⚠️ Nothing is stamped late — do not apply the real release's −0.900 s / −0.41 s fix (§3.6).
4. ⚠️ gello_joint_states/stamp_s is a monotonic clock, every other stamp_s is unix epoch (§3.3). t_rel_s is the common timebase.
5. Live camera rate 23.4 – 27.3 fps against a 30 fps container — the live videos play ~1.18× fast. Cause: software-GL rendering under CPU load. The recorder flags it per take in sim_meta.problems and sim_meta.achieved_fps_take. retimed_30hz/ is the fix; the largest grid-snap state-lookup error it introduces is 37.6 ms (§6).
6. cam1 / cam2 frame counts differ in 20 of 22 live takes (−9 … +5), because the two cameras render on independent workers. Equal in every retimed take. Map by nearest t_rel_s, never by index.
7. Seven takes contain more than one gripper closure (grip_cmd ≥ 0.7 rising edge): take_04, take_08, take_09, take_14, take_17, take_20, take_21 — a missed first grasp followed by a successful re-grasp (29 closures over 22 takes). All seven finish the task and all are included.
8. grip_pos while carrying the carrot ranges 0.565 – 0.913, not the real rig's tight 0.47 – 0.66 plateau: the sim carrot is tapered and the contact is soft, so how far the fingers close depends on where along the taper the grasp landed. Fully open is 0.000 – 0.003. Use sim_object_poses if you need to know what is held.
9. wrench is unvalidated against the real arm (§3.3).
10. synchronized is filled here and empty in the real release — the only structural schema difference (§3.5).
11. Zero NaN and zero Inf in any channel of any group of any take, both sets.
12. Cleanliness: exactly three files per take folder, no stray files, no hidden files, no sub-directories, in all 44 folders. The only absolute path is sim_meta.retimed.from in each retimed vectors.h5.
13. Duration outliers, not defects: take_09 (27.82 s) and take_08 (24.01 s) run about 1.6× the median take length (15.69 s) — slower demonstrations. Their data is clean.
14. First simulated session on this stack. These 22 takes are the first real GELLO→MuJoCo teleop takes ever recorded here; expect a less fluent motion style than the real-robot release's.

---

10. Reproducing these numbers

python3 scripts/dataset/make_carrot_raw_stats.py \
    --data <staging>/takes --out <staging>/dataset_stats.json \
    --dataset Bigenlight/carrot_in_pot_sim_raw
python3 scripts/dataset/make_carrot_raw_stats.py \
    --data <staging>/retimed_30hz --out <staging>/retimed_30hz/dataset_stats.json \
    --dataset Bigenlight/carrot_in_pot_sim_raw

Depth-free mode is auto-detected from the takes on disk (no depth.h5 present); --no-depth forces it and --depth demands the real release's four-file set. Requires h5py, numpy, cv2 and ffprobe on PATH.