DATA_DICTIONARY.md

Data Dictionary — Bigenlight/banana_in_pot_raw

Raw teleoperation recordings for the task "put the right banana in the pot", captured via a GELLO leader → UR7e follower setup with two RGB cameras.

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 datasets (banana_in_pot, banana_in_pot_ee) and convert_to_lerobot.py.

• Scale: 51 episodes (take_*) · 21,524 camera frames · 717.4 s (12 min) · ~745 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 51 files; no absolute-path leakage and no PII inside the H5.

---

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	Binary open/close command (grip_cmd) plus a continuous measured position (grip_pos), and the leader-side gello_grip.
Camera 1	Intel RealSense D435 — RGB only.
Camera 2	Intel RealSense D435if (D435 variant) — RGB only.
Video format	1280×720 (720p), 30 fps, MPEG-4 (mpeg4), color yuv420p. No depth / no IR recorded despite RealSense capability — color stream only. LeRobot copies re-encode to AV1; RAW keeps MPEG-4.
Viewpoints	Two fixed physical viewpoints (one tripod/third-person, one workspace). cam1↔cam2 order is meaningful — preserve it at deploy time.
Scene	Table with distractors (2 bananas, apple, carrots/peppers, watermelon slice), a silver pot, and an ArUco/AprilTag fiducial. Success = the right banana placed in the pot.

---

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

One file per take. Root has 9 groups, 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 (typical)	Native rate	What it is
cam1_frames	= cam1 frame count	30 Hz	Timestamp + frame index for each Camera-1 video frame
cam2_frames	= cam2 frame count	30 Hz	Timestamp + frame index for each Camera-2 video frame
command	~2× cam	~56–60 Hz	Commanded UR7e joint targets (the action), radians
ur_joint_states	~2× cam	~56–60 Hz	Measured UR7e joint state: position, velocity, effort
tcp_pose	~2× cam	~56–60 Hz	Measured tool-center-point pose (position + quaternion)
wrench	~2× cam	~56–60 Hz	6-axis force/torque at the TCP
gripper	~1.2× cam	~36–37 Hz	Gripper command, measured position, leader grip
gello_joint_states	= cam	~30 Hz	GELLO leader joint pos + vel (teleop only — not for inference)
synchronized	0 (EMPTY)	—	Intended fused/aligned table; empty in all 51 takes — ignore

Rate note: the spec quotes round figures (56 Hz / ~37 Hz). Measured average rates across the 51 takes are: command/ur/tcp/wrench median **59.6 Hz** (range 40–75), gripper median ~35.9 Hz (35.5–38.7), gello/cameras ~30 Hz. Rates vary per take because logging is timestamp-driven, not fixed-period — always resample using t_rel_s, never assume a fixed dt.

⚠️ 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", "cmd1", ...]' — NOT a native list/array. h5py returns it as a Python str. If a naive consumer does list(grp.attrs["columns"]) expecting a list, it iterates the string character-by-character and you get ['[', '"', 't', '_', 'r', 'e', 'l', ...] — 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.

cam1_frames — Camera-1 frame timeline

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

cam2_frames — Camera-2 frame timeline

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

command — commanded UR7e joint targets ➜ the ACTION

Channel	Unit	Meaning
t_rel_s	s	Timestamp
cmd1…cmd6	rad	Absolute target angle for UR7e joints 1–6 (base→wrist3)

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

Channel	Unit	Meaning
t_rel_s	s	Timestamp
q1…q6	rad	Measured joint angle, joints 1–6
qd1…qd6	rad·s⁻¹	Measured joint velocity, joints 1–6
eff1…eff6	N·m (motor effort/current-proxy)	Measured joint effort/torque, joints 1–6

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

Channel	Unit	Meaning
t_rel_s	s	Timestamp
x, y, z	m	TCP position in robot base frame
qx, qy, qz, qw	unit quaternion	TCP orientation (x,y,z,w order)

wrench — 6-axis force/torque at TCP

Channel	Unit	Meaning
t_rel_s	s	Timestamp
fx, fy, fz	N	Force along base/TCP x,y,z
tx, ty, tz	N·m	Torque about x,y,z

gripper — gripper signals

Channel	Unit	Meaning
t_rel_s	s	Timestamp
gello_grip	normalized	Leader (GELLO) grip trigger — teleop only
grip_cmd	normalized/binary open→close	Commanded gripper (part of the action). ⚠ all-NaN in one take — see anomalies
grip_pos	normalized 0–1	Measured gripper opening (part of observation.state)

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

Channel	Unit	Meaning
t_rel_s	s	Timestamp
q1…q6	rad	Leader joint angle
qd1…qd6	rad·s⁻¹	Leader joint velocity

synchronized — EMPTY in all takes (do not use)

Group exists with 56 declared channels (t_rel_s, t_wall, gello_q1..6, gello_qd1..6, gello_grip, cmd1..6, ur_q1..6, ur_qd1..6, ur_eff1..6, 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) but every dataset has shape (0,). Fusion was done downstream at conversion time, not stored here. Ignore it.

---

4. How to load (h5py)

import json, h5py, numpy as np

path = "Put_right_banana_in_the_pot/take_01_20260707_154624/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
    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)
    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 — align to the camera grid by nearest timestamp:
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].

5. How to convert to LeRobot

Use convert_to_lerobot.py (repo root). 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)
• gello_* streams are intentionally dropped (not observable at inference).
• The all-NaN grip_cmd take is repaired by forward/back fill (ffill_bfill).

lr_env/bin/python convert_to_lerobot.py --data Put_right_banana_in_the_pot \
    --out banana_in_pot_lerobot --repo-id Bigenlight/banana_in_pot

The extended-EE variant additionally exposes observation.tcp_pose (7) and observation.wrench (6).

---

6. Anomalies & data-quality notes

• All-NaN grip_cmd: take_01_20260707_180606 has gripper/grip_cmd = NaN for all 333 rows. grip_pos and gello_grip are fine. Converters must fill (fwd/back) or drop this channel for that take. No other NaNs anywhere in the dataset.
• Camera frame-count mismatch: in 21 / 51 takes cam1 and cam2 differ by ±1–2 frames (totals: 21,524 cam1 vs 21,515 cam2). Handle by nearest-timestamp mapping between the two camera clocks (as the converter does) — never assume cam1[k] and cam2[k] are simultaneous.
• synchronized/ empty in all 51 takes (see §3).
• Take-number gaps {9, 23, 30, 35} are by design (aborted/discarded takes); folder names span take_01…take_36 across four recording sessions (timestamps 15:46, 16:00, 17:37, 18:06 on 2026-07-07), and re-takes bring the folder count to exactly 51. Do not treat gaps as missing data.
• columns attribute is a JSON string, not a list — always json.loads it (see §2).
• Cleanliness: 0 stray files, 0 hidden files, no sub-directories, and no /home/ (absolute path) strings inside any of the 51 H5 files.

See dataset_stats.json for exact per-take frame counts, durations, and rates.