Humanoid Robot Training Data and Teleoperation
Humanoid data collection for robot learning, recorded by teleoperating the Unitree G1 in stationary and loco-manipulation tasks with grippers or dexterous and tactile hands. Every episode pairs camera views, depth, joint states and actions on one clock, delivered in LeRobot format as ready-made data or custom collection.
Embodiments and Robot Hands We Record With
Our Unitree teleoperation runs on the G1, a humanoid robot made by Unitree, with the Dex1 gripper, tactile hands or the Dex3, Inspire or BrainCo Revo 2 hands fitted to your task, so every episode matches the robot your policy will run on.
G1, stationary
The G1 holds its stance at a counter, table or shelf while the operator drives its arms, waist and hands.
G1, loco-manipulation
The operator drives walking and manipulation together, so one episode can cross a room and stock a shelf.
Dex1 gripper
Unitree's Dex1-1 gripper for pick-and-place, packing and tool handling, with one open and close action per hand.
Dex3, Inspire and BrainCo Revo 2 hands
Multi-finger hands for in-hand grasps and fine work, with contact recorded where the hand carries tactile sensors.
Two-arm tabletop rigs
Bimanual tabletop arms run on the same pipeline, and our teleoperation and capture services page covers them.
Learn more →Matched to your task
We match gripper or hand to the objects and grasps in your task list before the pilot.
Stationary and Loco-Manipulation Tasks
Humanoid teleoperation on the G1 runs in two modes. In stationary episodes the robot keeps its footing while the operator works its upper body. In loco-manipulation episodes walking and manipulation share one episode, so the policy learns to move to the work and then do it. What each mode records:
Kitchen, bedroom and staged retail tasks, plus your own.
Our ready-made data mostly covers kitchen, bedroom and staged retail tasks. Custom programs follow your task list, written as the objects, actions and success conditions your policy is evaluated on.
What Each Humanoid Episode Holds
Each episode is one synchronized recording of a task, from what the robot saw to what the operator commanded. Streams are agreed per program.
RGB video
Head camera views of the scene, the hands and the objects, the view a policy sees from the robot.
Depth
Metric per-frame depth, so positions and distances in each episode come out at true scale.
Wrist cameras
Left and right wrist views, the close-up perspective a robot has at execution time.
Joint states
Positions for the arms, waist and hands at every step, plus the legs in walking episodes.
Actions
The commands the operator sent at each step, the targets a policy learns to reproduce.
Tactile and force-torque
Contact force from tactile hands and force-torque readings where the hardware carries them. Our tactile manipulation data page covers that capture in depth.
Dexterous Hand Teleoperation Data
The end effector on the G1 determines what each episode records about the hands. We record with the Dex1 gripper, tactile hands and the Dex3, Inspire and BrainCo Revo 2 hands, and the one you choose sets how many hand values every step carries.
Dex1 gripper
Unitree lists the Dex1-1 gripper, which fits the G1, with an adjustable gripping force of 5 to 120 N and a 90 mm stroke. A gripper adds one open and close value per hand, which keeps the action space small.
Dex3
Unitree lists 7 degrees of freedom on the Dex3-1, three in the thumb and two each in the index and middle fingers, plus 33 pressure sensors in nine tactile arrays across the palm, finger pads and fingertips.
Inspire RH56
Inspire Robots lists its RH56DFX five-finger model at 6 degrees of freedom across 12 joints, with absolute position and force sensors that give real-time force feedback.
BrainCo Revo 2
The Revo 2 is the BrainCo hand we record with. BrainCo lists it at 11 degrees of freedom, 6 of them active, documents mounting it on the Unitree G1 and adds a fingertip tactile module on its Touch version.
Tactile hands
Hands fitted with tactile sensors add contact readings to every step, on the same clock as the cameras and joint states, with force-torque streams where the program calls for them.
What Finger Joints and Contact Readings Add
Finger joint states record how the hand closes, opens and regrasps an object through the episode, so a policy learns grasp shape along with arm motion. Contact readings show when each finger touches the object and how hard it presses, which helps a policy learn grip force on soft or slippery items and tell a secure grasp from one that is slipping.
How Hand Choice Changes the Action Space
A gripper adds one open and close value per hand to each action. A multi-finger hand adds a target for every actuated joint, which by the makers' figures means seven per hand on the Dex3-1 and six on the RH56DFX and the Revo 2. Two hands double those counts, and the state and action vectors in each delivered episode grow to match. We agree the end effector with you before the pilot, so every episode in a program shares one action space.
Tabletop and Walking Modes
Hand choice applies in both G1 modes. In stationary tabletop episodes the robot holds its stance at a counter or table while the hands do the work. In loco-manipulation episodes it walks to the work first, and the hand streams record alongside the leg joints.
Ready-Made G1 Sets
Our ready-made G1 sets cover kitchen, bedroom and staged retail task families, the staged retail episodes recorded with real teleoperation in a mock store set. Each set is a humanoid dataset of real G1 episodes, licensed non-exclusive, and a representative sample pack is available on request with the same modalities, calibration and delivery format. Teams post-training GR00T N1 on these episodes can see how each one maps to the model's inputs on our GR00T fine-tuning data for the Unitree G1 page.
Stationary Humanoid, Loco-Manipulation and Tabletop Arms Compared
Each setup records a different action space and suits different tasks.
| Setup | Stationary humanoid | Loco-manipulation | Bimanual tabletop arms |
|---|---|---|---|
| What is recorded | Head and wrist camera views, depth, upper-body and hand joint states and commanded actions, plus tactile and force-torque where fitted | Everything in a stationary episode, plus leg joint states and the robot's movement through the scene | Wrist and scene camera views, joint states for both arms, gripper or hand states and commanded actions |
| Action space | Arm, waist and hand joints, with the legs holding a stance | The whole body: legs, waist, arms and hands. Unitree lists the G1 at 23 joint degrees of freedom, and up to 43 in configurations with hands and optional waist and wrist joints | The joints of two fixed-base arms, plus each gripper or hand |
| Typical tasks | Counter and table work such as picking, placing, sorting and handing over objects | Tasks that span a room or aisle, such as fetching an item, tidying a bedroom or stocking a shelf | Two-handed tabletop work such as folding, packing, kitchen prep and assembly |
| What makes episodes harder to collect | Fine finger control on multi-finger hands and consistent scene resets between episodes | Balance and walking run alongside manipulation, scenes are larger to reset and each episode covers more ground | Timing two arms together, handovers between hands and deformable objects such as cloth |
Ready-Made Humanoid Data and Custom Collection
Ready-made data covers real-robot episodes on the G1, bimanual tabletop arms and UMI handheld capture, mostly kitchen, bedroom and staged retail tasks. If you need a Unitree G1 dataset before your own program is scoped, it gives your team real episodes to train and evaluate on. How the ready-made terms work:
Recorded to your tasks, on the terms you choose.
Custom collection runs on non-exclusive or exclusive terms. Exclusivity covers the data and its resale rights, separate from the robot that recorded it. Exclusive terms are priced above non-exclusive, and term and scope are agreed once the pilot is defined.
Delivery Formats and Episode Records
Episodes ship in LeRobot format, with MCAP, HDF5 and RLDS on request, so they load into the training stack your team already runs.
LeRobot by default
Camera streams, joint states and actions are packaged as LeRobot episodes, ready for policy training.
MCAP, HDF5 and RLDS on request
For teams whose loaders, replay tools or storage expect them, agreed with you in scoping.
Provenance and consent records
Every episode carries provenance and consent records, with signed participant and site agreements permitting commercial training and a chain of custody you can audit.
What Shapes the Engagement
We scope every humanoid program against these eight factors with you, then quote the program as a whole.
Embodiment and hands
The G1 or bimanual tabletop arms, with grippers or dexterous and tactile hands.
Tabletop or locomotion
Stationary work at a counter or table, or loco-manipulation between stations.
Tasks and environments
Task list, objects and settings, from kitchen, bedroom and staged retail families to tasks written for your evaluation.
Streams and format
Which camera, joint, tactile and force-torque streams each episode carries, and the delivery format.
Episode length
How much of a task one episode covers, from a single grasp to a multi-step routine.
Acceptance and QC criteria
What counts as a reject, such as a dropped object, a collision, a stream out of sync or an incomplete task.
Pilot size
How many episodes and task variations the pilot covers, so your team can train and evaluate on it.
Exclusivity and sublicense rights
Non-exclusive, sublicensable ready-made data, or custom collection on non-exclusive or exclusive terms.
Have a Humanoid Task to Record?
Send us the task, the environment and the hands your policy needs. We'll scope a pilot with the same streams, calibration and format you'd get at scale, so your team can check the episodes first.
Book a Strategy Session →Frequently Asked Questions on Humanoid Data Collection
Both. Stationary episodes keep the G1 at a counter, table or shelf and record its arms, waist and hands. Walking episodes add the legs, so the robot moves to the work and completes it within one episode. One program can mix both modes across its task list.
It depends on the objects and grasps in your task list. Parallel grippers suit pick-and-place, packing and tool handling, and they keep the action space small. Dexterous and tactile hands suit in-hand regrasping, fine assembly and soft objects, and they add finger joints and contact readings to each episode. We agree the end effector with you before the pilot.
Ready-made data is licensed non-exclusive, and you can sublicense it to your own client. Custom collection runs on non-exclusive or exclusive terms. Exclusivity covers the data and its resale rights, separate from the robot used to record it, and exclusive terms are priced above non-exclusive. Term and scope are agreed once the pilot is defined.
LeRobot by default, with MCAP, HDF5 and RLDS on request. Every episode carries its streams on one clock, with provenance and consent records alongside.
We write them with your team before the pilot runs. Each task gets a success condition and a list of rejects, such as a dropped object, a collision, a stream out of sync or an incomplete task. Episodes go through human verification against those criteria, and the same rules carry into scaled collection once your team has evaluated the pilot.
Real-robot episodes on the G1, bimanual tabletop arms and UMI handheld capture, mostly kitchen, bedroom and staged retail tasks, licensed non-exclusive and sublicensable to your own client.
Related Physical AI Solutions
Teleoperation and Capture Services for Robot Learning
Learn more →GR00T Fine-Tuning Data for the Unitree G1
Learn more →Tactile Manipulation Data: Tactile Sensing and Force Arrays for Robots
Learn more →Human Demonstrations for Robot Policies
Learn more →Physical AI Training Data: Multimodal Field Capture for Robot Learning
Learn more →Physical AI and Robotics Consulting, Development and Training Data
Learn more →World Model Training Data: Real Video for World Foundation Models
Learn more →Physical AI and Robotics Consulting
Strategy, data capture and model development for humanoid programs, from one team.
See physical AI services →AI Strategy Session
A 30-minute scoping call on your tasks, embodiment, hands and delivery format.
Book a session →Tactile Manipulation Data
Per-finger force arrays for contact-rich manipulation, synced to vision.
See tactile manipulation data →Scope Your Humanoid Robot Training Data
A 30-minute strategy call. Tell us the embodiment, hands and tasks, and we'll scope a humanoid data collection pilot with the streams, format and acceptance criteria your team needs.