Humanoid Teleoperation Hardware and Software: What the Rigs Contain and Cost

Four rig families and what each captures

A teleoperation rig is the equipment between the operator's body and the robot's joint targets. The documented rigs fall into four families: a headset that tracks head and hands and streams a stereo view back; leader arms that the operator moves while follower arms copy them; a wearable cockpit of exoskeleton arms, gloves and pedals; and full-body capture suits with instrumented gloves. The concept article on this site explains command rate, latency and retargeting; this one lists the parts. [1] [2] [3] [4]

How humanoid robot teleoperation works: rates, delays and retargeting

Headsets and web streaming: Open-TeleVision and Unitree's stack

Open-TeleVision, from UC San Diego and MIT, uses an Apple Vision Pro as the headset, streams the operator's head, wrist and hand poses to a web server built on Vuer, and returns a stereo image from a ZED Mini at 480 by 640 pixels per eye. The whole loop runs at 60 Hz. The robots were a Unitree H1 with Inspire hands, six actuated joints per hand, and a Fourier GR-1 with a one-joint gripper. Arm retargeting is closed-loop inverse kinematics on Pinocchio; hand retargeting uses dex-retargeting. The paper shows the same web interface on a Quest, a laptop and a phone, and reports an operator in Boston driving a robot in San Diego, without a latency figure. [1]

The repository adds the networking detail: an Ubuntu server on the same router as the headset, an HTTPS certificate from mkcert, port 8012, and ngrok for operation over the internet with Vision Pro or Quest 3. The README's example is a simulated robot. [5]

Unitree's official avp_teleoperate repository is the production version of the same idea. It lists the G1 in 29 and 23 joint configurations, the H1 with a four-joint arm, the H1_2 and H2 with seven-joint arms and the R1 with five or seven joints, plus the Dex1-1 gripper, Dex3-1 hand, Inspire hand and BrainCo hand. Headsets are Apple Vision Pro, PICO 4 Ultra Enterprise and Meta Quest 3 or 3S, in hand or controller input modes. The default control and recording frequency is 30 Hz. Vuer serves the operator view on port 8012 and a WebRTC image service on the robot's second computer serves video on port 60001. The hardware page requires the G1's developer computing unit and a Wi-Fi 6 router, recommends an x86-64 PC without stating a processor or GPU, and lists a 60 fps stereo camera with a 60 mm baseline or the built-in RealSense D435i. [6] [7]

Headset prices are official. Apple Vision Pro with the M5 chip starts at 3,499 US dollars. Meta listed the Quest 3S from 299.99 dollars and the 512 GB Quest 3 at 499.99 dollars in September 2024; a press report of April 2026 gives 599 dollars for the Quest 3 and 349 dollars for the 128 GB Quest 3S after Meta's increase of 19 April 2026, which Meta announced but whose figures could not be read on Meta's page. [8] [9] [10]

Leader-follower arms: ALOHA, ALOHA 2, Mobile ALOHA and GELLO

ALOHA, from Stanford, maps two WidowX leader arms joint by joint to two ViperX follower arms, with teleoperation and recording at 50 Hz and four Logitech C922x webcams at 480 by 640. The paper states that the whole system costs under 20,000 US dollars, with each ViperX around 5,600 dollars and the WidowX leader 3,300 dollars, and reports that cutting the rate from 50 Hz to 5 Hz lengthened teleoperation time by 62 percent. [2]

ALOHA 2 keeps the layout and changes the parts: a low-friction rail gripper on the leader that needs about ten times less force, 0.84 N against 14.68 N, a follower gripper force raised from 12.8 N to 27.9 N, passive gravity compensation in place of rubber bands, an aluminium extrusion frame and four RealSense D405 cameras at 848 by 480. It states no cost. Mobile ALOHA puts the arms on an AgileX Tracer base at 7,000 dollars with a 1.26 kWh battery and a laptop with an RTX 3070 Ti, for a stated 32,000 dollars including power and compute; the operator walks behind it and backdrives the wheels. [11] [12]

GELLO is the low end: a 3D-printed kinematic copy of the target arm fitted with Dynamixel XL-330 servos whose 12-bit encoders give 0.088 degrees of resolution, for a bill of materials under 300 dollars per device. The paper used UR5, xArm7 and Franka Panda; the software repository adds I2RT YAM, Franka FR3 and FER under an MIT licence. Neither states a control rate. [13] [14]

Whole-body cockpits: HOMIE and TWIST

HOMIE, from Shanghai AI Laboratory and CUHK, drives a Unitree G1 with Dex3-1 hands from a seated cockpit: two seven-joint exoskeleton arms with Dynamixel XL330-M288-T servos, gloves that read up to 15 finger joints with Hall-effect sensors and magnets, and a three-pedal foot unit for walking commands. The paper's cost table lists the exoskeleton at 430 dollars, each glove at 30 and the pedals at 20, about 510 dollars in 3D-printed PLA. The locomotion policy runs at 50 Hz on a Jetson Orin, the cockpit samples the exoskeleton at 260 Hz and the gloves at 300 Hz, arm targets reach the G1 at 10 Hz and are interpolated, and the measured communication latency over Wi-Fi is 16 ms. The hardware files are released under CC BY-NC-SA 4.0 with commercial use forbidden. [3] [15]

TWIST, from Stanford, is the capture-suit alternative on the same robot: an OptiTrack system at 120 Hz feeds a retargeting step and a 50 Hz policy over a 1 kHz PD loop, with inference on an RTX 4090. The paper puts the total delay at about 0.9 s, roughly 0.7 s of target generation and 0.2 s of inference. It states no cost; optical capture volumes are quoted, not listed. [16]

Suits and gloves: specifications public, prices mostly not

The suit figures matter for one reason: a 240 Hz suit with 20 ms latency is faster than any of the robot-side loops above, so in a suit-driven rig the delay comes from retargeting and policy inference, as TWIST's 0.9 s shows, not from the capture hardware. [4] [17] [18] [19] [16]

Software: LeRobot, dex-retargeting, Vuer and WebRTC

Hugging Face's LeRobot is the common open-source layer for leader-follower data collection. Its README lists SO-100, LeKiwi, Koch, HopeJR, OMX, EarthRover, Reachy 2, OpenARM, Unitree G1 and reBot B601 as supported hardware, gamepads, keyboards and phones as teleoperators, and GELLO, SpaceMouse and Meta Quest as community plugins, under Apache-2.0. dex-retargeting, from the AnyTeleop authors, provides the vector, position and DexPilot optimisers that turn a tracked human hand into robot hand joint targets, under an MIT licence; both Open-TeleVision and Unitree's stack use it. [20] [21] [1] [6]

On the streaming side, both headset stacks serve the operator view with Vuer over HTTPS on port 8012; Unitree adds a WebRTC image service on port 60001. The Open-TeleVision paper itself does not mention WebRTC. [1] [5] [6]

Rates and latencies in one table

What commercial humanoids say about teleoperation

1X's NEO page says the robot works autonomously by default, that an owner can schedule a 1X Expert to supervise it remotely at agreed times, and that the owner can pilot it through the mobile app and a VR device. The order page lists 20,000 dollars for early access or 499 dollars per month with a 200 dollar deposit and US deliveries from 2026. The headset model used by 1X operators is reported by the press, not stated by 1X. [22] [23]

Figure's Helix post of February 2025 says the model was trained on about 500 hours of teleoperated behaviours with its S1 controller at 200 Hz. The Helix 02 post of January 2026 states that the videos shown are fully autonomous and not teleoperated and cites over 1,000 hours of retargeted human motion data. Agility's FAQ says Digit runs autonomously in production and that teleoperation is used only during initial site setup to map a facility and for off-site development. [24] [25] [26]

For Tesla, only press reports exist: a 2024 job listing for operators wearing motion-capture suits and VR headsets at 25.25 to 48 dollars per hour, and a November 2025 report of helmet cameras, a backpack rig and a shift away from suits toward camera-only collection. No Tesla page states any of it, so these are recorded as press reports. [27] [28]