Robot teleoperation is the practice of a human operating a robot in real time through a control interface, and in robot learning it is the main way training demonstrations get recorded. This guide covers the five rig families and what they cost, what separates usable demonstration data from expensive noise, how much data published programs needed and how to calculate cost per usable episode rather than cost per hour.
Almost every capable manipulation policy shipped in the last three years was trained on data a person recorded by hand. The Open X-Embodiment dataset pooled more than one million real robot trajectories across 22 embodiments and 527 skills from 60 datasets and 34 labs. DROID contributed 76,000 demonstration trajectories, 350 hours of interaction across 564 scenes and 86 tasks, gathered by 50 human collectors at 13 institutions over 12 months.
What Is Robot Teleoperation?
Robot teleoperation means a human controls a robot remotely in real time, with the operator’s movements or inputs mapped onto the robot’s joints or end effector. It serves two separate purposes.
The first is operation itself, as in surgery, subsea inspection or warehouse exception handling, where a person does a job through a machine. The second is data collection, where the recording is the point of the session. Every teleoperated episode produces synchronized robot actions and sensor observations that become training data for imitation learning and vision-language-action models.
This guide addresses the second. The distinction matters commercially, because a rig built for one purpose suits the other less well. Operation rewards latency and situational awareness. Collection rewards throughput, consistency and clean synchronization.
Teleoperation Systems: The Five Rig Families
Hardware choice is driven by task family first and budget second. The table below sets out what each approach costs and where it earns its place.
| Rig family | Reference example | Hardware cost | Suits |
|---|---|---|---|
| Bimanual leader-follower | ALOHA, Mobile ALOHA | About USD 18,000, or 20,000 with camera add-ons | Fine two-handed manipulation: folding, threading, insertion |
| Low-cost kinematic leader | GELLO | Under USD 300 per device in parts, follower arm excluded | Single-arm collection where you already own the robot |
| VR headset and controllers | Meta Quest, Open-TeleVision, TWIST2 | Roughly USD 500 to 3,500 plus the robot | Humanoids, mobile manipulation, active head cameras |
| Handheld robot-free capture | Universal Manipulation Interface (UMI) | Under USD 1,500 per device | In-the-wild scene diversity, and collection before hardware arrives |
| Dexterous hand and glove systems | DexPilot, Shadow Teleoperation System | From about USD 100,000 to over 400,000 | Multi-finger dexterity and contact-rich specialist work |
Two findings from the literature are worth carrying into a hardware decision. ALOHA reaches 80 to 90 percent task success when paired with action chunking, at hardware under 20,000 US dollars, which is why it appears behind most bimanual VLA papers. UMI removes the robot from the collection loop entirely, so trajectories are recorded in any environment and retargeted to the robot action space during training.
ALOHA and Mobile ALOHA
The ALOHA robot comes from a 2023 paper by Tony Z. Zhao, Vikash Kumar, Sergey Levine and Chelsea Finn, and its project page describes it as a low-cost, open-source teleoperation system. The operator backdrives two small WidowX leader arms, and two larger ViperX follower arms copy them through direct joint-space mapping. Teleoperation and recording both run at 50Hz. Each step stores the leader joint positions as the action, with the follower joint positions and four 480×640 RGB camera feeds as the observation, two cameras fixed on the frame and two on the robot wrists. That is the episode layout the paper’s Action Chunking with Transformers (ACT) policy learns from.
Mobile ALOHA, from Zipeng Fu, Tony Z. Zhao and Chelsea Finn at Stanford, puts the arms on an AgileX Tracer base. The operator is tethered to the base at the waist and backdrives its low-friction wheels, so one person drives the arms and the base together, and each episode adds the base’s linear and angular velocity to the joint positions of all four arms. The paper reports a 32,000 US dollar build budget, including onboard power and compute. With 50 demonstrations per task, co-training on existing static ALOHA datasets raised success rates by up to 90 percent on tasks such as sautéing and serving a piece of shrimp, storing heavy pots in a two-door wall cabinet and calling and entering an elevator.
ALOHA 2, from a Google DeepMind team whose authors include Chelsea Finn and Tony Z. Zhao (arXiv 2405.02292), keeps the leader-follower design and improves performance, ergonomics and durability over the original. It adds low-friction rail grippers on the leader and follower arms, a passive gravity compensation mechanism built from off-the-shelf parts, a simpler frame that keeps the camera mounts rigid and four smaller Intel RealSense D405 cameras with depth in place of the original webcams. The team open-sourced all hardware designs with a tutorial and released a MuJoCo model of the workcell with system identification for simulation work.
GELLO Leader-Follower Arms
GELLO comes from a paper by Philipp Wu and colleagues at UC Berkeley (arXiv 2309.13037). Each device is a small teleoperation robot arm, a scaled copy of the target arm with a kinematically equivalent structure, built from 3D-printed parts and off-the-shelf DYNAMIXEL XL330 motors. The paper puts the bill of materials for each GELLO device under 300 US dollars, and the team built versions for the UR5, the xArm7 and the Franka Panda.
In GELLO teleoperation, the software reads joint angles from the leader’s motors as the operator moves it and sends them straight to the robot as joint commands, so the action stored at each step is a joint-position target. In the paper’s user study, 12 participants ran five two-arm tasks on a pair of UR robots and completed them more reliably and faster with GELLO than with VR controllers or a SpaceMouse.
What Separates Good Teleoperation Data From Expensive Noise?
Control Rate and Fidelity
Demonstrations should be recorded at the policy’s control rate, commonly 30 to 60 Hz. Below that you get less signal per minute of operator time. The interface also needs to transmit operator intent accurately, since imprecise control produces trajectories that teach the policy imprecision.
Consistency Across Sessions
The same operator should produce similar demonstrations week to week. High variance between sessions limits what a policy can extract, and differences in teleoperator skill add noise that is hard to spot by eye.
Coverage Rather Than Polish
DROID was designed around scene diversity rather than repetition, moving robots freely between sessions instead of returning to the same tabletop. Policies co-trained with it improved in-distribution success by 22 percentage points and out-of-distribution success by 17 points against an Open X-Embodiment baseline. Breadth of scene, viewpoint, object and lighting is what buys generalisation.
Curation Beats Collection
More data and better data are different targets. Published curation work reports that Dexora improved policy performance while discarding 80 percent of its training demonstrations using a kinematic screen. Methods including CUPID and Demo-SCORE use influence functions and policy rollouts to identify which individual demonstrations help closed-loop success.
Budget for a QA and curation stage from the start. A smaller, properly filtered dataset often outperforms the raw capture it came from.
How Much Teleoperation Data Do You Need?
These reference points come from published programs.
| Program | Scale | Collection footprint |
|---|---|---|
| Single-task fine-tune | Low hundreds of demonstrations | One rig, days to weeks |
| Fourier ActionNet | 30,000 bimanual trajectories, about 140 hours | VR-based egocentric teleoperation on humanoids |
| DROID | 76,000 trajectories, 350 hours, 564 scenes, 86 tasks | 18 Franka arms, 50 collectors, 13 institutions, 12 months |
| RT-1 | 130,000 episodes across 700+ tasks | 13 robots, 17 months |
| AgiBot World | Over 1 million trajectories, 217 tasks, 106 scenes | 100+ identical humanoid platforms |
| Open X-Embodiment | Over 1 million trajectories, 527 skills | 22 embodiments pooled from 60 datasets, 34 labs |
Read that table as a ladder of ambition, from a single-task fine-tune on hardware you already own to a foundation-model corpus built across many institutions.
Cost per Hour, and the Number That Matters More
All-in teleoperation collection has been benchmarked at 28 to 60 US dollars per operator hour once wages, supervision, facility overhead and hardware amortization are counted. That figure is the starting input.
Two multipliers sit on top of it. Teleoperated collection runs close to one hour of skilled operator time for every hour of usable robot data, a constraint AgiBot researchers describe in their work on scalable collection, and only part of each session survives QA. The number to manage combines both:
Cost per usable episode = (operator hours × all-in hourly rate) ÷ usable episode yield
Hardware is the smallest line in most budgets: a 20,000 dollar bimanual station running two shifts amortizes to a small share of a year of operator wages. The recurring costs sit in operator time, rig maintenance and QA review, and each rig family spreads them differently.
Ask any collection partner for yield and rejection rate alongside the hourly figure. A lower rate with a 55 percent yield loses to a higher rate with a 90 percent yield, and only one of those two numbers usually appears on a quote.
Methods That Lower Cost per Usable Episode
Synthesis from a single demonstration is the sharpest lever. The Real2Render2Real work reported policies trained on data generated from one human demonstration matching policies trained on 150 teleoperated demonstrations, measured across 1,050 physical robot evaluations. Robot-free capture with UMI, human video co-training and simulation each shift the ratio further.
Robot Teleoperation: Build In-House or Contract the Collection?
Build or Contract – both routes work. The decision turns on how long you will be collecting and how specialised the task is.
| Factor | Build in-house | Contract a collection partner |
|---|---|---|
| Time to first episode | Weeks, including rig build and operator training | Days, on an existing floor |
| Cost shape | Capital plus ongoing wages, technician time and supervision | Variable, priced per episode or per operator hour |
| Task specificity | Suits proprietary tasks on your own hardware | Suits common manipulation skills and scene breadth |
| Scene diversity | Limited to your facility unless you travel | Multiple sites and environments already available |
| Break-even point | Favourable once collection runs continuously for many months | Favourable for bounded programs and for testing whether data helps at all |
A common pattern splits the two. Teams keep depth in-house on the handful of tasks that define their product and contract out breadth, the wide range of scenes, objects and viewpoints that buys generalization. That split also gives you a yield benchmark, because you can compare your own usable episodes per operator hour against a partner running the same task.
Building a floor front-loads hardware, rig build, operator training and pipeline engineering, costs paid once rather than per episode. A first in-house run of 1,000 demonstrations carries all of them while later runs on the same rig reuse them, so most of the robot data collection cost you pay in year one sits in that gap. Blend your own cost across the first two or three programs and set it against the partner’s rate.
How to Set Up a Teleoperation Data Program?
- Pick the rig from the task. Fine bimanual work needs leader-follower. Humanoid and mobile manipulation needs VR with active head cameras. Scene breadth points to handheld capture.
- Fix the control rate and camera set before episode one. Changing either mid-program splits your dataset into incompatible halves.
- Write the QA rubric first. Define what a rejected episode looks like, who reviews and at what sampling rate.
- Standardize the format. LeRobot and HDF5 are the common targets, and the LeRobot dataset format now packs many episodes into shared Parquet and MP4 shards rather than one file per episode.
- Record provenance and consent. Operator identity, location, date, task label and rights terms, captured per episode rather than reconstructed later.
- Measure yield weekly. Usable episodes per operator hour is your unit economic. Track it so you can tell whether operator training is working.
- Budget a curation stage. Assume a meaningful share of raw capture will be filtered or down-weighted, and plan the spend accordingly.
Scoping a teleoperation program or comparing build against buy? NeuralChainAI runs robot data collection services and embodied AI training data programs, with the rig selection, QA thresholds and format decisions settled before the first episode is recorded. Bring the task list and the target policy; you get back a costed collection plan with yield assumptions written down.
Teams weighing a data program alongside a wider deployment can also start from physical AI and robotics consulting, where hardware, integration and data strategy are scoped together.