README.md

metadata

pretty_name: 'rhoyn Teleoperation Sample: Unitree G1, OpenArm and reBot'
license: cc-by-4.0
task_categories:
  - robotics
size_categories:
  - 10K<n<100K
tags:
  - LeRobot
  - humanoid
  - unitree-g1
  - openarm
  - rebot
  - teleoperation
  - whole-body
  - bimanual
  - manipulation
  - depth
  - lidar
configs:
  - config_name: wipes
    data_files: wipes/data/*/*.parquet
  - config_name: socks
    data_files: socks/data/*/*.parquet
  - config_name: eggs
    data_files: eggs/data/*/*.parquet
  - config_name: openarm_eggs
    data_files: openarm/eggs/data/*/*.parquet
  - config_name: openarm_connect_four
    data_files: openarm/connect_four/data/*/*.parquet
  - config_name: openarm_jenga
    data_files: openarm/jenga/data/*/*.parquet
  - config_name: openarm_puzzle
    data_files: openarm/puzzle/data/*/*.parquet
  - config_name: openarm_socks
    data_files: openarm/socks/data/*/*.parquet
  - config_name: openarm_wipes
    data_files: openarm/wipes/data/*/*.parquet
  - config_name: rebot_eggs
    data_files: rebot/eggs/data/*/*.parquet
  - config_name: rebot_connect_four
    data_files: rebot/connect_four/data/*/*.parquet
  - config_name: rebot_jenga
    data_files: rebot/jenga/data/*/*.parquet
  - config_name: rebot_puzzle
    data_files: rebot/puzzle/data/*/*.parquet
  - config_name: rebot_socks
    data_files: rebot/socks/data/*/*.parquet
  - config_name: rebot_wipes
    data_files: rebot/wipes/data/*/*.parquet

rhoyn Teleoperation Sample: Unitree G1, OpenArm and reBot

Teleoperation data from rhoyn, recorded with our own low-latency VR teleoperation stack: the operator wears a PICO 4 Ultra headset and the robot's arms and grippers follow their hands precisely. The sample covers three robots: a Unitree G1 whole-body humanoid, a bimanual OpenArm (two 7-DoF arms) and a bimanual reBot B601-RS (two 6-DoF arms). Each episode records every sensor on the robot alongside a standard LeRobot v3.0 dataset. This is a free sample of the format; rhoyn collects data to your spec. See also rhoyn.com/data.

Unitree G1

wipes

socks

eggs

Each G1 preview tiles the nine camera streams (color and stereo IR) above the lidar point cloud. The loops are 6 s; the full-length previews are in previews/.

OpenArm

eggs	connect four	jenga
puzzle	socks	wipes

Each OpenArm preview is about 8 s: three moments of the episode joined with short crossfades. The top half shows a scene camera beside the D435F's lossless depth (warm is near, cool is far, black is no reading); below are every robot camera in color and the left image of its infrared stereo pair. The scene camera (a ZED) is only there for the previews and is not part of the dataset.

reBot

eggs	connect four	jenga
puzzle	socks	wipes

Each reBot preview is about 8 s of the episode, in the same layout as the OpenArm previews: a scene camera beside the D435F's lossless depth on top, then every robot camera in color and the left image of its infrared stereo pair. The scene camera (a ZED) is only there for the previews and is not part of the dataset.

At a glance

	Unitree G1	OpenArm	reBot
Robot	Unitree G1 Edu+, 29 DoF whole-body humanoid, Dex1-1 grippers	OpenArm, two 7-DoF arms (bimanual), Dex1-1 grippers	reBot B601-RS, two 6-DoF arms (bimanual) with grippers
Teleoperation	rhoyn VR teleop, PICO 4 Ultra headset and controllers	rhoyn VR teleop, PICO 4 Ultra headset and controllers	rhoyn VR teleop, PICO headset and controllers
Episodes	3, one per task · 12,633 frames · 7 min	6, one per task · 20,841 frames · 11.6 min	6, one per task · 23,543 frames · 13.1 min
Rate	30 Hz rows; depth at 15 fps; lidar, IMUs and VR packets at native rates	30 Hz rows; depth at 15 fps; VR packets at native rate	30 Hz rows; depth at 15 fps; VR packets at native rate
Cameras	9 streams: RealSense D435i on the head + two wrist D405s, each color + stereo IR	9 streams: RealSense D435F on top of the column + two wrist D405s, each color + stereo IR	9 streams: RealSense D435F on a camera post + two wrist D405s, each color + stereo IR
Depth	Lossless 16-bit from all three cameras	Lossless 16-bit from all three cameras	Lossless 16-bit from all three cameras
Lidar	Livox MID-360 points, raw packets and IMU	None	None
IMUs	Pelvis, chest, D435i	None	None
Format	LeRobotDataset v3.0 per task folder, plus sidecar streams	LeRobotDataset v3.0 per task folder, plus sidecar streams	LeRobotDataset v3.0 per task folder, plus sidecar streams
Size	2.26 GB	4.26 GB	5.16 GB
License	CC BY 4.0	CC BY 4.0	CC BY 4.0

Tasks

Unitree G1

Folder	Task	Duration	Frames	Preview
wipes/	Pull a wet wipe from the pack and wipe the table	97 s	2,921	previews/wipes.mp4
socks/	Pick up the two socks and pair them together	92 s	2,747	previews/socks.mp4
eggs/	Crack eggs into the pan and scramble them	232 s	6,965	previews/eggs.mp4

Three more G1 tasks (connect four, jenga and a puzzle) are in the full six-task G1 download at rhoyn.com/data.

Per G1 episode: 9 synchronized camera streams (a RealSense D435i on the head + two wrist cameras, color and stereo IR), lossless 16-bit depth from all three cameras, a Livox MID-360 lidar point cloud, three IMUs, the full joint state and torque pipeline, and the raw VR teleoperation packets.

OpenArm

Folder	Task	Duration	Frames	Preview
openarm/eggs/	Put butter in the pan, crack the egg into it and scramble it	192 s	5,755	previews/openarm/eggs.mp4
openarm/connect_four/	Pick up a disc, pass it to the other gripper and drop it into the connect four board	80 s	2,388	previews/openarm/connect_four.mp4
openarm/jenga/	Push a block out of the Jenga tower, pass it to the other gripper and place it on top	94 s	2,825	previews/openarm/jenga.mp4
openarm/puzzle/	Pick up the puzzle pieces and place them	104 s	3,127	previews/openarm/puzzle.mp4
openarm/socks/	Pick up the two socks and pair them together	108 s	3,228	previews/openarm/socks.mp4
openarm/wipes/	Open the pack, pull out a wet wipe and wipe the table	117 s	3,518	previews/openarm/wipes.mp4

Per OpenArm episode: 9 synchronized camera streams (a RealSense D435F on top of the column + two wrist cameras, color and stereo IR), lossless 16-bit depth from all three cameras, the full joint state and torque pipeline (PD gains, feedforward and velocity targets, requested/capped/commanded torque, torque budget and limiter state, measured joint torque, motor temperature), and the raw PICO VR teleoperation packets. There is no lidar and there are no IMUs on the OpenArm.

reBot

Folder	Task	Duration	Frames	Preview
rebot/eggs/	Put butter in the pan, crack the egg into it and scramble it	261 s	7,821	previews/rebot/eggs.mp4
rebot/connect_four/	Pick up a disc, pass it to the other gripper and drop it into the connect four board	73 s	2,177	previews/rebot/connect_four.mp4
rebot/jenga/	Push a block out of the Jenga tower, pass it to the other gripper and place it on top	138 s	4,135	previews/rebot/jenga.mp4
rebot/puzzle/	Pick up the puzzle pieces and place them	118 s	3,527	previews/rebot/puzzle.mp4
rebot/socks/	Pick up the two socks and pair them together	91 s	2,725	previews/rebot/socks.mp4
rebot/wipes/	Open the pack, pull out a wet wipe and wipe the table	105 s	3,158	previews/rebot/wipes.mp4

Per reBot episode: 9 synchronized camera streams (a RealSense D435F on a camera post + two wrist cameras, color and stereo IR), lossless 16-bit depth from all three cameras, the full joint state and torque pipeline (PD gains, feedforward and velocity targets, requested/capped/commanded torque, torque budget and limiter state, measured joint torque, motor temperature), and the raw PICO VR teleoperation packets. There is no lidar and there are no IMUs on the reBot.

Each folder holds one episode. The task column below is the episode's language instruction, exactly as stored in meta/tasks.parquet.

Loading

Each task folder is its own LeRobot dataset. Install LeRobot with its dataset extra (tested with lerobot 0.6.1):

pip install "lerobot[dataset]"

Then download one folder and point LeRobot at it:

from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

# Unitree G1
path = snapshot_download("rhoyn-robotics/data-sample", repo_type="dataset", allow_patterns="wipes/*")
ds = LeRobotDataset("rhoyn-robotics/data-sample", root=f"{path}/wipes")

# OpenArm
path = snapshot_download("rhoyn-robotics/data-sample", repo_type="dataset", allow_patterns="openarm/wipes/*")
ds = LeRobotDataset("rhoyn-robotics/data-sample", root=f"{path}/openarm/wipes")

# reBot
path = snapshot_download("rhoyn-robotics/data-sample", repo_type="dataset", allow_patterns="rebot/wipes/*")
ds = LeRobotDataset("rhoyn-robotics/data-sample", root=f"{path}/rebot/wipes")

Or from the command line:

hf download rhoyn-robotics/data-sample --repo-type dataset --include "wipes/*" --local-dir rhoyn-sample
hf download rhoyn-robotics/data-sample --repo-type dataset --include "openarm/wipes/*" --local-dir rhoyn-sample
hf download rhoyn-robotics/data-sample --repo-type dataset --include "rebot/wipes/*" --local-dir rhoyn-sample

Loading by repo id alone (LeRobotDataset("rhoyn-robotics/data-sample")) is not supported, because the repo root is not itself a LeRobot dataset. Use the root= form above.

Videos are AV1-encoded. Decoding needs an AV1-capable decoder (dav1d or libaom) in your ffmpeg / PyAV / torchcodec build. Recent pip wheels of PyAV and torchcodec include one.

The per-frame data can also be read directly with pandas or pyarrow from <task>/data/chunk-000/*.parquet (G1), openarm/<task>/data/chunk-000/*.parquet (OpenArm) or rebot/<task>/data/chunk-000/*.parquet (reBot).

Control and actions

The operator wears a PICO VR headset and controllers.

Unitree G1. The robot is driven by two controllers at once:

• Arms and grippers (operator). The 14 arm joints and 2 grippers follow the operator's tracked wrist poses (retargeted to joint targets) and the grip triggers. Only these joints receive feedforward torque (action.tau_ff).
• Legs and waist (learned controller). The 15 leg and waist joint targets come from a GR00T whole-body-control locomotion policy, which switches between balance and walking modes. The operator steers it with the controller sticks; those commands are recorded in action.teleop as vx, vy, yaw_rate and height_command. For these joints action.tau_ff and action.dq_target are always 0.

For policy learning on the G1, the operator-driven action is the arm and gripper part of action (or the raw wrist poses in action.teleop), plus the locomotion commands in action.teleop. The leg and waist entries of action are included for completeness and for whole-body training.

OpenArm. All 14 arm joints and both grippers follow the operator: the arms track the operator's wrist poses (retargeted to joint targets) and the grippers follow the grip triggers. action is the full operator-driven action. action.teleop has the same 26-element layout as on the G1; its locomotion fields (vx, vy, yaw_rate, height_command) are not used on the OpenArm.

reBot. The same as the OpenArm, with 12 arm joints: both arms track the operator's wrist poses and both grippers follow the grip triggers. action is the full operator-driven action, and the locomotion fields of action.teleop are not used.

Reference

Contents of each episode folder

<task>/  (G1)        openarm/<task>/  (OpenArm)        rebot/<task>/  (reBot)
├── meta/            info.json, stats.json, tasks.parquet, episodes/, plus sidecar metadata
│                    (calibration.json, depth.json, lidar.json, d435_imu.json, pico.json)
├── data/            per-frame parquet (state, actions, IMU, torques, teleop, sync info)
├── videos/          9 AV1 video streams, one per camera / IR imager
├── depth/           lossless 16-bit depth (FFV1 .mov) for 3 cameras, at 15 fps
├── lidar/           Livox MID-360 points, raw packets, and lidar IMU        (G1 only)
├── d435_imu/        D435i IMU                                               (G1 only)
└── pico/            raw Pico VR teleop controller packets

The sidecar folders (depth/, lidar/, d435_imu/, pico/) are not part of the LeRobot format. LeRobot tools such as aggregate_datasets ignore them, so read them directly using the paths and metadata in meta/.

OpenArm and reBot episodes have no lidar/ or d435_imu/ folder and no meta/lidar.json. Their meta/d435_imu.json and the lidar and d435_imu entries (null) in meta/calibration.json come from the shared recording stack and point to no data.

Cameras (videos/)

Key	Resolution	Description
observation.images.d435	848×480	RealSense D435i on the head (G1), D435F on top of the column, tilted 20° (OpenArm), or D435F on a camera post, tilted 20° (reBot); color
observation.images.d435_ir1, _ir2	848×480	D435i (G1) or D435F (OpenArm, reBot), stereo IR
observation.images.d405_left, d405_right	424×240	Wrist RealSense D405, color
observation.images.d405_{left,right}_ir1, _ir2	424×240	Wrist D405, stereo IR

The camera keys and resolutions are the same on all three robots.

IR streams are monochrome but stored as 3-channel video (the gray value is replicated), so they load as (3, H, W) like the color streams.

On all three robots the right wrist camera is mounted upside down and its frames are rotated upright at recording time (frames_rotated_upright: true in meta/calibration.json). The left wrist camera's frames are stored as captured.

Per-frame data, Unitree G1 (data/)

Key	Dim	Description
observation.state	60	29 joint positions, 29 joint velocities, 2 gripper apertures
observation.effort	31	Measured joint torques
observation.motor_temp	29	Motor temperatures
observation.imu.pelvis, observation.imu.chest	10 each	Quaternion, gyro, accelerometer
action	31	29 joint position targets, 2 gripper aperture targets
action.kp, action.kd, action.tau_ff	29 each	Low-level PD gains, feedforward torque
action.dq_target †	29	Velocity targets
action.tau_requested, action.torque_cap, action.tau_commanded †	31 each	Torque pipeline: requested, capped, commanded
observation.torque_budget, action.torque_allowed, action.torque_limited †	16 each	Arm torque-safety limiter state
action.teleop	26	Operator VR wrist poses, locomotion commands (vx, vy, yaw rate, height), triggers, grips, sticks
action.teleop.buttons	8	VR controller buttons
observation.teleop_active	1	Whether teleop was engaged (always 1 in this sample)
observation.frame_counter, sensor_ts_us, frame_offset_us	12 each	Per-stream camera sync info
observation.row_time_us, observation.control_offset_us	1 each	Row and control-loop timing

Full joint names and per-element units for every vector are in meta/info.json.

† Not present in eggs/. The eggs episode was recorded on 2026-10-05, before the arm torque-safety limiter was added to the recording stack, so it has 31 features instead of 38. wipes/ and socks/ share an identical schema and can be merged with LeRobot's aggregate_datasets; eggs/ cannot be merged with them without dropping those 7 keys.

Per-frame data, OpenArm (openarm/<task>/data/)

Key	Dim	Description
observation.state	30	14 joint positions, 14 joint velocities, 2 gripper apertures
observation.effort	16	Measured joint torques (14 arm joints, 2 grippers)
observation.motor_temp	14	Motor temperatures
action	16	14 joint position targets, 2 gripper aperture targets
action.kp, action.kd, action.tau_ff	14 each	Low-level PD gains, feedforward torque
action.dq_target	14	Velocity targets
action.tau_requested, action.torque_cap, action.tau_commanded	16 each	Torque pipeline: requested, capped, commanded
observation.torque_budget, action.torque_allowed, action.torque_limited	16 each	Torque-safety limiter state
action.teleop	26	Operator VR wrist poses, triggers, grips, sticks (locomotion fields unused)
action.teleop.buttons	8	VR controller buttons
observation.teleop_active	1	Whether teleop was engaged (always 1 in this sample)
observation.frame_counter, sensor_ts_us, frame_offset_us	12 each	Per-stream camera sync info
observation.row_time_us, observation.control_offset_us	1 each	Row and control-loop timing

All six OpenArm episodes share an identical schema (robot_type: openarm_v2_2x7dof_dex1_1) and can be merged with LeRobot's aggregate_datasets. They cannot be merged with the G1 episodes, whose joint vectors differ. Full joint names and per-element units are in meta/info.json.

Per-frame data, reBot (rebot/<task>/data/)

Key	Dim	Description
observation.state	26	12 joint positions, 12 joint velocities, 2 gripper apertures
observation.effort	14	Measured joint torques (12 arm joints, 2 grippers)
observation.motor_temp	12	Motor temperatures
action	14	12 joint position targets, 2 gripper aperture targets
action.kp, action.kd, action.tau_ff	12 each	Low-level PD gains, feedforward torque
action.dq_target	12	Velocity targets
action.tau_requested, action.torque_cap, action.tau_commanded	14 each	Torque pipeline: requested, capped, commanded
observation.torque_budget, action.torque_allowed, action.torque_limited	14 each	Torque-safety limiter state
action.teleop	26	Operator VR wrist poses, triggers, grips, sticks (locomotion fields unused)
action.teleop.buttons	8	VR controller buttons
observation.teleop_active	1	Whether teleop was engaged (always 1 in this sample)
observation.frame_counter, sensor_ts_us, frame_offset_us	12 each	Per-stream camera sync info
observation.row_time_us, observation.control_offset_us	1 each	Row and control-loop timing

All six reBot episodes share an identical schema (robot_type: seeed_rebot_b601_rs_2x6dof) and can be merged with LeRobot's aggregate_datasets. They cannot be merged with the G1 or OpenArm episodes, whose joint vectors differ. Full joint names and per-element units are in meta/info.json.

Note on NaNs: observation.frame_counter, observation.sensor_ts_us and observation.frame_offset_us have three depth columns (d435_depth, d405_left_depth, d405_right_depth). On the G1 and OpenArm they are NaN on every row. On the reBot they are filled on the even rows, which carry a depth frame, and NaN on the odd rows. This is expected: depth is recorded at 15 fps in depth/, and meta/depth.json maps each depth frame to its data row.

Synchronization

All streams are timestamped against one host clock. The data rows are a fixed 30 Hz grid (observation.row_time_us), and each camera contributes the frame nearest to each row. The cameras run freely and are not hardware-triggered, so each camera frame is within about ±16.7 ms (half a frame) of its row time, and D435i/D435F and wrist frames in the same row can be up to about 30 ms apart (G1, OpenArm). On the reBot they are typically 12–20 ms apart, and up to about 120 ms apart on the few rows around a dropped frame. The exact capture time and offset of every stream's frame are recorded per row in observation.sensor_ts_us and observation.frame_offset_us.

Lidar points, IMU samples (G1) and Pico packets (all three robots) keep their native rates and carry their own timestamps. meta/lidar.json, meta/d435_imu.json and meta/pico.json describe the clocks.

Depth

Depth is lossless 16-bit (FFV1, gray16le), 424×240 at 15 fps, for all three cameras, on all three robots. Depth frame i belongs to data row 2i. Convert to meters with meters = raw * depth_scale (0.001 for the D435i and D435F, 0.0001 for the D405s, given per stream in meta/depth.json).

Two raw values mean "no valid depth" and should be masked:

• 0: no depth (no stereo match).
• 65535: saturated or out of range. This mostly happens on the wrist D405s when the scene is beyond their short working range.

The stats in meta/depth.json exclude both values.

Calibration

Camera intrinsics (all color, IR and depth streams) are RealSense factory calibration, read from each device at recording time.

Camera-to-robot (and, on the G1, lidar-to-robot) extrinsics come from the robot model, not from a hand-eye calibration. On the G1, the D435i and lidar poses come from the Unitree G1 URDF; the wrist cameras' poses come from the CAD model of their 3D-printed mounts. On the OpenArm, the D435F pose comes from the CAD model of its 20° column mount and the wrist cameras' poses from their Dex1-1 brackets. On the reBot, the D435F pose comes from the same 20° mount on the B601 camera post and the wrist cameras' poses from the B601 wrist mount model. Every transform carries a source field saying this. These poses are suitable for coarse fusion. If you need precise depth-to-robot or lidar-to-camera alignment, plan to refine them.

License and attribution

This sample is released under CC BY 4.0. You can use it for anything, commercial work included, with credit to rhoyn.

@misc{rhoyn2026g1sample,
  title        = {rhoyn Teleoperation Sample: Unitree G1, OpenArm and reBot},
  author       = {{rhoyn}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/rhoyn-robotics/data-sample}}
}

Privacy

Episodes were recorded in indoor home and office settings by the rhoyn team. Anyone who appears in the camera streams or previews, on any robot, is a rhoyn team member who agreed to their publication. The Pico packets contain controller poses relative to the operator's head heading, with no identifying information.

Custom data collection

rhoyn collects teleoperation data for your robot and your tasks, objects and scenes, on our own low-latency VR teleop stack, and delivers it in this format (LeRobotDataset v3.0 plus the extra sensor streams).

• Robots: Unitree G1, OpenArm, reBot and YAM.
• Free pilot: 50 episodes of your task. Train on them and keep them.
• Licensing: exclusive or non-exclusive, or full ownership transfer, matched to the project.

Contact hello@rhoyn.com or book a 30-minute call. Hardware and format details are at rhoyn.com/data.