Robots · Humanoids

Humanoid robot teleoperation, model by model

Humanoids learn from people. Here is how each model is driven remotely, what data it produces, who can actually buy one, and what it takes to run operators on it every day.

Humanoid robot teleoperation, model by model
Illustration

Updated

Every humanoid that does useful work today learned it from human demonstrations, and most still need a person to take over when they get stuck. The hardware differs a lot from model to model. The teleoperation problem is the same: map a human body onto a robot body, fast enough and cleanly enough that the result can be trained on.

How humanoids are teleoperated

VR headsets with hand tracking

An Apple Vision Pro or Meta Quest streams the operator's head and hand poses; the robot's cameras stream back to the headset. Open stacks such as OpenTeleVision and Unitree's xr_teleoperate made this the default for upper-body manipulation data.

Motion-capture suits and exoskeletons

Full-body suits or arm exoskeletons capture the whole body, including the legs and torso. They cost more and take longer to set up, but they are the way to collect loco-manipulation and whole-body data.

Joysticks plus a learned controller

For walking, most teams let the robot's own locomotion policy keep balance and give the operator a joystick for direction and speed, while the arms follow VR or a leader device.

Supervision and takeover

On robots that already run a policy, the operator watches several units and takes over only when one asks for help. Each takeover becomes a correction episode for the next training run.

What makes humanoid data good

Consistent operators. The same people, trained on the same protocol, give the policy one way to do the task rather than ten.
Clean episodes: a defined start pose, a clear success or failure, and a reset between attempts.
Complete streams: every camera, joint states, hand states and the operator's commands, time-aligned.
Honest labels. A failed grasp labeled as a failure is useful; a failed grasp hidden in a success is poison.

Our guide on human-in-the-loop fallback covers the takeover side, and the data collection guide covers formats and quality checks.

Several humanoid makers have sent us robots to test our operating protocols. For your program, pilots are trained on your unit and your tasks, and the same group stays with you.

Pick your humanoid

Each page below covers one model: specs and SDK, how it is teleoperated today, public datasets, whether you can actually get one, and how we staff it.

ModelTypeBest forDriven with
Unitree G1UnitreeCompact research humanoid (1.32 m)Manipulation and loco-manipulation datasets, research fleetsunitree_sdk2 (C++ / Python), unitree_ros2, xr_teleoperate (Vision Pro, PICO, Quest 3), LeRobot unitree_g1
Tesla OptimusTeslaGeneral-purpose humanoid (not for sale)Tesla internal factory use onlyMotion-capture suit (data collection), VR headset teleop, Helmet and backpack camera rigs, No public SDK
1X NEO1X TechnologiesHome humanoid with remote Expert ModeHousehold chores, with scheduled remote expertsVoice and 1X mobile app, Expert Mode (remote 1X operator), Quest 3 headset + controllers (1X Experts), Redwood AI on-board policy
Unitree H2UnitreeFull-size humanoid (1.82 m, 31 DoF)Human-height manipulation and whole-body dataunitree_sdk2 (EDU), xr_teleoperate v1.6 (H2, 7-DoF arms), NVIDIA Isaac TeleOp + GR00T (H2 Plus), unitree_mujoco (C++)
Unitree R1UnitreeLow-cost compact humanoid (1.23 m)Multi-robot tabletop data fleets, teaching labsunitree_sdk2 (EDU), xr_teleoperate v1.6 (R1_A5, R1_A7), Vision Pro, PICO, Quest 3, LeRobot v3.0 datasets
Unitree H1UnitreeFull-size research humanoid (H1 and H1-2)Whole-body teleop research, full-size loco-manipulationunitree_sdk2 / unitree_ros2, xr_teleoperate (H1 4-DoF, H1_2 7-DoF), Open-TeleVision (Vision Pro, Quest 3), OmniH2O / H2O (non-commercial)
Figure 03Figure AIGeneral-purpose humanoid (commercial deployments only)Figure's own commercial and home programsHelix / Helix 02 VLA (on-board), Teleoperation equipment (Figure operators), Human video and motion data, Figure Index app (human data)
Boston Dynamics AtlasBoston DynamicsElectric industrial humanoidAutomotive and industrial material handlingVR teleop on BD's MPC controller, Tablet steering interface, Autonomous mode (LBM policies), Orbit fleet software
AgiBot A2AgiBotFull-size service and performance humanoidReception, guiding, performances, recorded motionAgiBot VR Teleoperation Kit (A2 only), AimMaster record and playback, Handheld remote controller, Swarm and dance control software
Fourier GR-1FourierFull-size research humanoidVR teleop data collection and GR00T researchFourier teleoperation (F.A.R.T.S.), Apple Vision Pro / Meta Quest 3, GRx SDK and fourier-grx-client, ROS 2 on Ubuntu 20.04 host
Agility DigitAgility RoboticsBipedal warehouse and logistics humanoidTote handling and material moves in warehousesAgility Arc fleet cloud, Warehouse and MES integration, Swappable ISO-flange end-effectors, No public SDK or teleop stack
UBTech Walker S2UBTech RoboticsIndustrial humanoid with autonomous battery swapFactory material handling around the clockSDK via UBTech partner training, WalkerS2-Model URDF / USD, GHRC 2026 Isaac Sim baseline (LeRobot v3.0), Gigabit Ethernet / USB 3.0
Booster T1 (and Booster T2)Booster RoboticsMid-size research humanoid (T1, now T2)RoboCup teams, locomotion research, custom teleop buildsBooster Robotics SDK (C++ / Python), Booster Studio, booster_gym (Isaac Gym), booster_train / booster_deploy
AgiBot X2AgiBotCompact interaction and performance humanoidPerformances, interaction, motion data on UltraHandheld controller and mobile app, Secondary development (X2 Ultra only), AIMA Open Platform, Custom VR or mocap teleop
Pollen Robotics Reachy 2Pollen Robotics (Hugging Face)Wheeled bimanual humanoid (open source)VR teleop datasets in LeRobot, mobile manipulation researchReachy2Teleoperation VR app (Windows), LeRobot reachy2 robot and teleoperator, pollen_data_acquisition_server, Python SDK (reachy2-sdk)
EngineAI PM01EngineAIMid-size open-SDK humanoid (1.40 m, 24 DoF)Locomotion and motion-tracking research, custom teleop buildsEngineAI native SDK (ROS 2 Humble / Jazzy), engineai_ros2_workspace (Cyclone DDS), engineai_rl_lab (Isaac Lab), MuJoCo sim2sim
LimX Dynamics OliLimX DynamicsFull-size humanoid (165 cm, 31 DoF body)Whole-body motion data, mocap-driven teleop, researchRemote Control and High-Level Motion APIs, limxsdk-lowlevel (C++ / Python), GMR retargeting (Xsens, OptiTrack, PICO), Studio Package (data collection, optional)
Booster K1Booster RoboticsKid-size humanoid (95 cm, 22 DoF)RoboCup, locomotion and motion data, low-cost fleetsBooster Robotics SDK (C++ / Python), Booster Studio and Agents, booster_train (Isaac Lab), booster_deploy (MuJoCo, Webots)
AgiBot Genie G1AgiBotWheeled dual-arm data-collection humanoidTeleop data collection and GO-1 fine-tuningVR or motion-capture teleop (built in), Genie Studio, AgiBot World / LeRobot format, GO-1 foundation model
Rainbow Robotics RB-Y1Rainbow RoboticsWheeled bimanual humanoid (mobile manipulator)Leader-arm bimanual datasets, mobile manipulation researchrby1-sdk (Python / C++), ROS 2, Leader Arm and Stand Leader Arm (14 DoF), Tele-Op by VR (option)

Frequently asked questions

How are humanoid robots teleoperated?
Most often with a VR headset that tracks the operator's head and hands and maps them onto the robot's arms and hands, while the robot's own controller handles balance. Full-body work uses motion-capture suits or exoskeletons, and walking is usually steered with a joystick.
Which humanoid robots can you actually buy?
As of October 2026 the humanoids you can order are mostly Chinese research platforms: Unitree (G1, H1, H2, R1), Booster, AgiBot X2, EngineAI and LimX, sold directly or through resellers. Agility Digit and the Rainbow Robotics RB-Y1 are sold through sales teams, 1X NEO is on preorder, and Tesla Optimus and Figure 03 are not for sale. Each model page states its availability with a source.
How much humanoid teleoperation data do you need to train a policy?
It depends on the task and the model. Fine-tuning a pretrained vision-language-action model on one task often starts at a few hundred demonstrations; general skills need far more. A one-week trial on your robot gives you the rate per operator hour to plan from.
Can remote operators take over a humanoid when it gets stuck?
Yes. The robot asks for help, an operator takes control through the same interface used for data collection, finishes the task and hands control back. The takeover is recorded as a correction episode.