Robot Teleoperation: Methods, Hardware, and Best Practices
Robot teleoperation compared: VR headsets, leader-follower arms, data gloves and SpaceMouse, with live US prices for the hardware, how to plan a latency budget, and how to set up and scale data collection.
This guide covers how to control robots remotely, from VR headsets to leader-follower arms, with practical guidance on latency, data quality and scaling operations. Hardware prices below come live from our US store.
What Is Robot Teleoperation?
Robot teleoperation is the real-time remote control of a robot by a human operator. The operator sends continuous or near-continuous commands (position, velocity or force targets) through an input device, and the robot executes them as faithfully as its hardware and control system allow.
Teleoperation is far older than robot learning. It began with manipulators for handling radioactive material. What changed in the 2020s is the main use: teleoperation now mostly generates training data for robot learning. When a human teleoperates a robot through a task, the recorded sensor data and action commands become demonstrations that can train an imitation learning policy to perform the same task autonomously.
This dual use, as both a remote-control method and a data source, makes teleoperation one of the most important capabilities in modern robotics. The method you choose determines the quality of the data you collect, the tasks you can demonstrate and the throughput of your collection pipeline.
Applications of Teleoperation
- Demonstration data collection: the dominant application today. Operators teleoperate robots through manipulation tasks and record datasets of (observation, action) pairs. These train policies with behaviour cloning, ACT, Diffusion Policy or VLA models. See how to collect robot training data for the full pipeline.
- Remote inspection and maintenance: operating robots in hazardous places such as nuclear facilities, the deep sea, space or disaster zones, where people can't safely go. These jobs tolerate more latency and need less data quality.
- Telesurgery: surgical systems such as Intuitive's da Vinci are the best-known commercial teleoperation platforms. Surgery demands the lowest latency and highest precision of any application.
- Remote operation of deployed robots: operators in one place supervise or take over robots working somewhere else. See the remote fleet management guide for the monitoring, connectivity and recovery side.
Teleoperation Methods Compared
VR Headset Teleoperation
The operator wears a standalone VR headset and uses hand controllers or hand tracking to set the desired end-effector pose in Cartesian space. The robot's inverse kinematics (IK) solver computes the joint angles needed to reach it.
How It Works in Detail
The headset streams controller or hand poses (position and orientation) over the network to a workstation running the IK server. The server converts each end-effector target into joint positions, often with a solver built on MuJoCo, PyBullet or the robot vendor's own kinematics. It then sends those positions to the robot's controller over its native interface (USB, Ethernet or ROS 2). The operator watches a live camera feed from the robot in the headset and closes the loop visually.
| VR hardware | Our price | Notes |
|---|---|---|
| Meta Quest 3S | $450.00 | Standalone headset with 6-DOF inside-out tracking and Touch Plus controllers |
| Unitree VR Remote Operation Kit (PICO) | $2,400.00 | VR teleoperation accessory for the Unitree G1 ecosystem. Confirm compatibility with your robot before ordering |
| Unitree G1 Teleop Kit (no robot) | $15,000.00 | G1 teleoperation kit, sold without the robot |
Pros
- Lowest hardware cost: the headset is the only extra hardware
- Intuitive for operators with VR experience, because hand position maps naturally to robot position
- Works with any robot arm that accepts Cartesian position commands, and with humanoids through retargeting
- One operator can control both arms of a bimanual robot with two controllers
- Portable: the headset is not mechanically coupled to the robot, so the operator can sit in another room
Cons
- More latency than joint-space methods, from wireless transmission and IK computation
- No proprioceptive feedback: the operator can't feel arm resistance, joint limits or contact forces
- VR discomfort limits session length for many operators
- Lower precision on contact-rich tasks: without force feedback, operators tend to overshoot insertions and alignments
- IK singularities cause jerky motion near full extension or wrist-lock configurations
Leader-Follower Arms
The operator physically holds and moves a lightweight "leader" arm, and a "follower" arm tracks the leader's joint positions in real time. This is the approach of ALOHA, Mobile ALOHA and many current data-collection setups.
How It Works in Detail
In the original ALOHA paper, the operator backdrives a smaller WidowX arm (the leader) whose joints are synchronised with a larger ViperX 6-DoF arm (the follower), both from the same manufacturer. The authors chose this "direct joint-space mapping" instead of mapping a VR controller pose through IK. ACT policies on ALOHA predict joint positions at 50 Hz (ALOHA project page). Because the mapping is joint to joint, no IK runs in the loop.
The leader can be smaller and lighter than the follower. What matters is that the kinematics match, so the operator's sense of where their hand is maps directly to the follower's workspace. Gravity compensation on the leader makes it easier to hold during long sessions.
| Leader-follower hardware | Our price | Notes |
|---|---|---|
| RC Aloha Leader | $2,500.00 | ALOHA leader arm for dual-arm teleoperation |
| i2RT YAM Leader | $2,999.00 | Leader arm with bilateral force feedback from the follower |
| i2RT BIG YAM Leader | $3,449.00 | Leader arm for the BIG YAM |
Pros
- Low latency: joint-to-joint copying needs no IK, so operators feel directly connected to the robot
- Proprioceptive feedback: the operator feels the leader arm's position, which maps to the follower
- Clean joint-space data, with no IK artifacts or singularity issues
- Natural, fast operator movements
- Well proven: ACT was introduced on ALOHA's leader-follower hardware, and LeRobot records with leader-follower arm pairs
Cons
- Extra hardware cost for the leader arm (see the table above)
- The operator must be next to the leader arm
- Kinematic differences between leader and follower (link lengths, joint limits) cause workspace mismatches
- The leader's gravity compensation needs tuning for operator comfort, and poor tuning causes fatigue
For two-arm rigs, see the bimanual teleoperation hardware setup and the ALOHA guide.
Exoskeleton and Data Gloves
The operator wears a data glove or hand exoskeleton that tracks finger joint angles and, on some models, provides haptic feedback. Finger positions are retargeted to a dexterous robot hand.
| Glove hardware | Our price | Notes |
|---|---|---|
| Wuji Glove | $4,500.00 | Tactile and pose-tracking data glove |
| UDEXREAL UDCAP Data Glove | $3,500.00 | Finger-tracking data glove |
| RC G1 Tactile Glove | $3,625.00 | Robotics Center tactile glove. Check the product page for compatible hands |
MANUS gloves are available by quote. Compare data gloves side by side.
Pros
- Essential for dexterous manipulation: no other method gives finger-level control
- Haptic feedback, on gloves that have it, lets operators feel grasped objects
- A natural interface that uses the operator's existing fine motor skills
Cons
- Physically demanding over long sessions
- Needs calibration for each operator's hand
- Only useful with a dexterous robot hand, which is a significant additional cost
- Slower throughput, because dexterous tasks are inherently slower
See the glove-based dexterous teleoperation guide for glove selection, calibration and retargeting.
SpaceMouse, Joystick and Gamepad
The operator uses a 3Dconnexion SpaceMouse (six axes) or a gamepad to command end-effector or joint velocities. Each axis maps to one degree of freedom.
Pros
- The cheapest option, using off-the-shelf input devices
- No VR discomfort and little physical exertion, so operators can work long sessions
- Good for mobile robot navigation, camera positioning and simple pick-and-place
Cons
- Gamepads lack enough analog axes for a 6-DOF end effector. A SpaceMouse has six, but it still only commands velocity
- An unnatural interface: operators must mentally map device axes to robot axes
- Velocity control produces less natural trajectories than position control
- Hard to coordinate two arms or a gripper at the same time
See the SpaceMouse teleoperation guide for drivers, axis mapping and dead-zone tuning.
| Method | Control mapping | Operator feedback | Main strength | Best for |
|---|---|---|---|---|
| VR headset | Cartesian pose through IK | Visual only | Low cost, portable, works with humanoids | Budget teams, remote operators, humanoids |
| Leader-follower | Joint to joint | Proprioceptive (force feedback on some leaders) | Clean, fast demonstrations | Arm and bimanual data collection |
| Gloves / exoskeleton | Finger joints, retargeted | Visual, haptic on some gloves | Finger-level control | Dexterous hands |
| SpaceMouse / gamepad | End-effector or joint velocity | Visual only | Cheap, comfortable | Navigation, simple tasks, prototyping |
Latency: The Core Challenge
Latency is the delay between the operator's command and the robot's response. It is the single most important performance metric of a teleoperation system: high latency degrades operator performance, lowers data quality and makes contact-rich tasks impractical.
Latency Budget Breakdown
Total latency is the sum of several pipeline stages. Measure each one on your own system:
- Input device to computer: wired USB devices are the fastest. Wi-Fi headsets add network delay. Avoid Bluetooth input devices for teleoperation.
- Computation: joint-space leader-follower is close to a direct copy. Cartesian IK with collision checking costs more.
- Computer to robot actuators: direct USB or EtherCAT links are faster than routing commands through ROS 2 topics. The ROS 2 overhead depends on QoS settings and the network.
- Actuator response: the motor controller's settling time, which is longer for large motions than for small ones.
- Video back to the operator: camera exposure, encoding, transmission and display. This is often the largest share in remote or VR setups.
Planning Targets
Use these as planning targets, not guarantees:
- Around 10 ms or less: feels immediate. Achievable with a local leader-follower setup.
- Tens of milliseconds: feels responsive and is adequate for most manipulation. Realistic for VR on a good local network.
- Around 40 to 100 ms: a noticeable delay. Operators slow down and move more deliberately, and data from fast tasks suffers. Still usable for slow manipulation and navigation.
- Over 100 ms: significantly impairs operators, who fall back to a move-and-wait strategy. Acceptable only for non-time-critical work such as remote inspection, or with shared autonomy.
How to Minimize Latency
- Use wired connections (USB, Ethernet) instead of Wi-Fi wherever possible
- Run the control loop on a real-time kernel (Ubuntu with PREEMPT_RT) on dedicated CPU cores
- Cut ROS 2 middleware overhead with intra-process communication for co-located nodes
- Talk to motors directly (for example, Dynamixel Protocol 2.0 sync read/write) instead of through ROS 2 topics for the leader-follower link
- Profile the pipeline end to end and measure actual latency. A single misconfigured network buffer can add more delay than every other stage combined
Teleoperation for Data Collection vs. Real Deployment
Teleoperation requirements differ sharply depending on whether you are collecting training data or doing productive work.
Data Collection Teleoperation
- Priority: data quality and throughput. Every episode should be a clean, successful demonstration.
- Latency tolerance: moderate. The operator works in a controlled lab with the robot in direct view.
- Session duration: long sessions, often several per day, so operator comfort and fatigue management matter.
- Failure handling: discard failed demonstrations. A failed episode costs only the operator's time: fail fast, reset, repeat.
- Network: local network only. The latency and reliability penalties of running data collection over the internet are rarely worth it.
Deployment Teleoperation
- Priority: reliability and safety. The robot is doing real work, so failures have consequences: damaged goods, safety incidents, downtime.
- Latency tolerance: low, but wide-area links make high latency unavoidable. Manage it with predictive displays, shared autonomy and move-and-wait interfaces.
- Session duration: full shifts, which need ergonomic workstations, operator rotation and fatigue monitoring.
- Failure handling: safe fallback behaviours (stop, retreat to a safe pose, alert a supervisor). The robot should never enter an unrecoverable state.
- Network: usually remote, over WAN or the internet. Needs redundant connectivity, graceful degradation on packet loss, encryption and authentication. See the remote fleet management guide.
Robotics Center Teleoperation and Data Services
- Hardware: leader arms, VR headsets and gloves in the tables above, with complete rigs priced in the teleoperation rig cost breakdown.
- Data platform: our data platform is LeRobot and Hugging Face native, with quality dashboards for collected episodes.
- Managed collection: teleoperation data services with trained operators, if you'd rather buy demonstrations than build the station.
How to Set Up Your First Teleoperation Session
These steps assume you have a robot arm and want to start collecting demonstration data.
Step 1: Choose your teleoperation method
For most teams starting out on arms, leader-follower gives the cleanest data if a matching leader arm exists for your robot. If budget is tight, or you are teleoperating a humanoid, start with VR. If your task needs dexterous fingers, you need gloves. Use the comparison above to match the method to your task.
Step 2: Set up the hardware
Mount the robot arm on a heavy, rigid table, bolted or clamped down. Position two or three cameras: one overhead or third-person view, one on the wrist, and optionally a side view. Connect everything to one workstation (Ubuntu with a current ROS 2 LTS release). Check that every sensor stream is publishing and timestamped. The camera setup guide covers placement, synchronisation and calibration, and the teleoperation lab setup guide covers the room.
Step 3: Install teleoperation software
For low-cost arms such as the SO-101 and Koch, LeRobot provides leader-follower teleoperation and recording. For ALOHA-style ViperX/WidowX rigs, use the ALOHA codebase or LeRobot. For OpenArm, follow the OpenArm learning path. For industrial arms, use the vendor's external-control or streaming interface with your own IK bridge.
Step 4: Calibrate and test
Run teleoperation in test mode, without recording. Check that the robot follows your commands smoothly, that the cameras cover the whole task workspace, that latency feels acceptable and that joint limits don't interfere with the task. Adjust camera positions, tune IK parameters and test the emergency stop.
Step 5: Define your task protocol
Before collecting any data, write down the task goal (specific and unambiguous), the starting configuration (object positions, arm pose), the expected approach and the success and failure criteria. A written protocol keeps data consistent across operators and sessions.
Step 6: Collect a small test batch
Record a few dozen demonstrations, train a simple policy (for example ACT with default hyperparameters) and evaluate it on the real robot. If it performs poorly, review the demonstrations for inconsistency before blaming model capacity. Then refine the protocol and collect again. The OpenArm learning path walks through this end to end.
Scaling Teleoperation Operations
Moving from one research station to a multi-station collection operation takes standardisation, quality control and operator management.
Station Standardization
Every station must produce physically equivalent data. That means the same arm models and firmware versions, the same camera models in the same positions (use a mounting jig), controlled lighting instead of ambient light, and software versions frozen at the start of the campaign.
Operator Training Pipeline
New operators need structured training before they collect production data:
- Free-form practice on the teleoperation interface.
- Practice on the specific task with immediate feedback.
- A calibration batch evaluated against your quality criteria.
- A production session shadowing an experienced operator.
Only operators who meet the quality bar on their calibration batch join the production rotation.
Quality Control System
Use three layers of quality control:
- Real time: automated checks during collection, such as episode length bounds, joint-limit violations and camera frame drops.
- Post-episode: human or automated review soon after collection, covering success and strategy consistency.
- Pre-training: dataset-level analysis of distribution coverage, outliers and consistency between operators.
Reject episodes that fail any layer, and track the rejection rate. A rising rate points to a problem with the protocol or with operator training.
FAQ
Which teleoperation method gives the best training data? For arm manipulation, leader-follower arms are the most widely used choice, because the joint-to-joint mapping avoids IK artifacts. VR is the cheapest and the usual choice for humanoids. Gloves are required for dexterous hands.
How much latency is acceptable for robot teleoperation? For local data collection, aim for tens of milliseconds or less. Around 40 to 100 ms is noticeable, and operators slow down. Above 100 ms you need move-and-wait or shared autonomy. Measure your own pipeline end to end, including the video path back to the operator.
What does a teleoperation rig cost? It depends on the method. Live prices for leader arms, headsets and gloves are in the tables above, and the teleoperation rig cost guide adds up complete rigs.
Can I try a teleoperation setup before buying? Yes. Contact us to arrange a visit to our San Francisco lab, or ask about leasing.