How to Become a Robot Teleoperator: The Job, Skills and Data

Robot teleoperator was not a job title five years ago. It is one now, because modern robot policies train on human demonstrations and somebody has to record them. This page covers the work hour to hour, what separates a demonstration that helps a model from one that quietly poisons it, and which pay figures survive being checked.

Two paths run through it. Get an arm, install lerobot, record episodes, train on them. Or work through a platform that already owns the robots and the customers. Most people good at the second spent a few weekends on the first.

What a robot teleoperator actually does

Teleoperation means driving a robot you are not touching. In robot learning it has one purpose: producing demonstrations. You perform the task through the robot's body, the software records every joint angle and camera frame at a fixed rate, and that recording becomes one episode. Enough episodes make a policy, a network mapping camera frames and joint states to the next actions. That is imitation learning, and you supply the imitation.

The common rig on cheap arms is leader-follower: you move an unpowered copy of the arm by hand and the powered one mirrors your joint angles. That is much of why the SO-100 became the default teaching arm. Other rigs change the job: VR controllers, a 3D mouse, a phone held in the air as in RoboTurk, or a mocap suit for a humanoid.

Listings mix these four freely, so read the physical requirements before the title. Fortune reported in August 2024 that Tesla's Data Collection Operator role paid 25.25 to 48 USD per hour and required walking more than seven hours a day in a motion-capture suit and a VR headset. Fortune gives the height window as 5 ft 7 to 5 ft 11; the posting itself says 5 ft 7 to 6 ft. Read the posting, not the coverage: it also names Palo Alto, three shift options, up to 30 lbs of equipment and daily written reports.

The fallback category is growing on its own. 1X is taking pre-orders for NEO, a home robot with teleoperation built in, at 20,000 USD or 499 USD a month. Owners schedule when an operator may take over and name the task. Engadget reports that the company can blur people out of the operator's view, that owners can designate no-go zones, and that operators cannot take control without the owner's approval.

An hour on the clock

The arithmetic surprises people. DROID holds 76,000 trajectories totalling 350 hours, roughly 17 seconds of robot motion per trajectory. lerobot's recorder defaults to 60 seconds per episode and 60 seconds of reset. The demonstration is the short part; half your clock, by default, is putting the cube back.

So an hour is ten minutes of setup, then a loop until the counter hits its target. The commands below are the lerobot CLI at 0.6.1, released 3 August 2026, which needs Python 3.12 or newer. Version matters here: the module behind lerobot-record moved to lerobot.scripts.lerobot_record in 0.4.0, so a python -m invocation from an older tutorial no longer resolves, and running a trained policy moved into a separate lerobot-rollout in 0.6.0.

1. 1
   Install the right extras, then find the serial ports

   Since 0.6.0 a bare install pulls no dataset, hardware or visualisation dependencies, so lerobot-record does not run. The extra that maps to the recording scripts is core_scripts; STS3215 servos need the Feetech SDK on top. Then find the ports: the tool asks you to unplug the USB cable and prints the one that disappeared.

   bash

   # NOT enough on 0.6.x: pip install lerobot
   pip install 'lerobot[core_scripts,feetech]'
   
   lerobot-find-port
   lerobot-find-cameras

2. 2
   Calibrate both arms

   Put every joint roughly mid-range first. The calibration file is keyed by the id you pass here, so reuse that id in every later command.

   bash

   lerobot-calibrate \
       --robot.type=so101_follower \
       --robot.port=/dev/ttyACM0 \
       --robot.id=my_follower_arm
   
   lerobot-calibrate \
       --teleop.type=so101_leader \
       --teleop.port=/dev/ttyACM1 \
       --teleop.id=my_leader_arm

3. 3
   Dry run the teleop link before recording

   Drive the task five times without recording. Check that the follower tracks without lag, that both views cover the workspace, and that you can finish looking only at the feeds.

   bash

   lerobot-teleoperate \
       --robot.type=so101_follower \
       --robot.port=/dev/ttyACM0 \
       --robot.id=my_follower_arm \
       --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
       --teleop.type=so101_leader \
       --teleop.port=/dev/ttyACM1 \
       --teleop.id=my_leader_arm \
       --display_data=true

4. 4
   Record the block of episodes

   One task string per dataset, phrased as an instruction starting with a verb. Right arrow saves the episode and moves on, left arrow re-records the last one, Escape stops and encodes. Over a laggy SSH or VNC link use the single-byte equivalents n, r and q: arrow keys arrive as multi-byte escape sequences the link can split. The dataset is pushed to your Hugging Face account by default, so --dataset.push_to_hub=false is what keeps a client's kitchen off the public Hub.

   bash

   lerobot-record \
       --robot.type=so101_follower \
       --robot.port=/dev/ttyACM0 \
       --robot.id=my_follower_arm \
       --robot.cameras="{ top: {type: opencv, index_or_path: 1, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30} }" \
       --teleop.type=so101_leader \
       --teleop.port=/dev/ttyACM1 \
       --teleop.id=my_leader_arm \
       --dataset.repo_id=${HF_USER}/so101_pickplace \
       --dataset.num_episodes=30 \
       --dataset.single_task="Put the red brick in the bowl" \
       --dataset.streaming_encoding=true \
       --display_data=true

5. 5
   Reset the scene deliberately, and resume rather than restart

   New object position every episode, but fixed camera, lighting and strategy. The lerobot guide suggests at least 50 episodes, about 10 per location. Resuming the next day has three traps. --dataset.num_episodes now means additional episodes, not the new total; --dataset.root must be set or the resume fails; and the directory is not the name you typed, because on creation lerobot-record stamps _YYYYMMDD_HHMMSS onto the repo id unless you passed --dataset.no_stamp=true.

   bash

   # the first run stamped the repo id, so look up what it actually created
   ls ~/.cache/huggingface/lerobot/${HF_USER}/
   
   # resume it with the same robot and teleop flags as above;
   # num_episodes is now ADDITIONAL episodes, not the new total
   lerobot-record \
       --dataset.repo_id=${HF_USER}/so101_pickplace_20260823_141500 \
       --dataset.root=~/.cache/huggingface/lerobot/${HF_USER}/so101_pickplace_20260823_141500 \
       --dataset.num_episodes=20 \
       --resume=true

What separates a useful operator from an expensive one

Nobody watches you work, and the feedback loop is long: you record for two hours, someone trains for four, and the result tells you something about your recordings you can no longer reconstruct.

1. One strategy per task, across every episode. Approaching from the left in episode 3 and from above in episode 4 does not give the model two options, it gives it a coin flip at the same observation.
2. Constant pacing. Policies using action chunking predict a block of future actions from one observation, so a demonstration that starts slow and ends panicked teaches an inconsistent mapping.
3. Vary the world, not the method. Object position and clutter change between episodes; camera placement, lighting, grip strategy and speed do not.
4. Drive from the camera feed. The lerobot guide states the rule of thumb: you should be able to do the task yourself looking only at the camera images.
5. Stop when you get tired. Fatigue shows up in the last third of a session as systematic bias, not useful noise.

There is published evidence that operator skill changes the outcome. The robomimic study released each task in two variants so the effect could be measured.

Note what that implies about hiring. A team does not need six identical experts; it needs to know which operator produced which episode, so a bad batch can be traced and dropped.

What happens to the hour you just recorded

Your session lands on disk as a LeRobot dataset: per-frame observations and actions, camera streams encoded as video, plus metadata for the features and frame rate. That shared format is why an episode recorded on an SO-100 in your kitchen trains in somebody else's pipeline without conversion.

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("your-username/so101_pickplace")

print("episodes:", ds.num_episodes)
print("frames:  ", ds.num_frames)
print("fps:     ", ds.fps)
print("features:", list(ds.features))

# frames / (episodes * fps) is your mean episode length in seconds.
# Far below your task time means episodes are being cut short.

From there the dataset goes into a training run and comes out as a checkpoint. Two details decide whether it even starts. A LeRobot v3.0 dataset crashes the GR00T loader and has to be converted down to v2.1. And the episode count, which varies by model.

Those are floors, not targets: they say when a fine-tune has enough to bite, not when the policy is good. If a job dies on the loader or never leaves the queue, dataset rejected as v3 and training job stuck queued come first.

Does hobbyist data help? The SmolVLA team curated 487 community datasets focused on the SO-100 arm, standardised at 30 fps, around 10 million frames, and pre-trained on them. Success on SO-100 went from 51.7 percent without that pre-training to 78.3 percent with it. Their own description of the corpus is varied lighting, suboptimal demonstrations, unconventional objects, heterogeneous control schemes. They kept the mess and cleaned only the labels, rewriting each instruction to under 30 characters starting with an action verb.

Skills and hardware you actually need

The skill list is short and none of it is a degree. Steady hands matter more than programming, and patience more than either, because the fiftieth repetition has to look like the fifth.

• Hand-eye coordination from a 2D camera feed rather than direct sight. This separates people fast, and it is trainable in a few hours.
• Tolerance for repetition. The job is closer to a machinist's than a programmer's.
• Basic terminal literacy: install a package, pass flags, read an error.
• Enough mechanical sympathy to notice a servo running hot before it strips.
• A stable connection if the robot is not in the room with you.

That cap has a number. The control loop runs from 20 ms per action step for ACT to 485 ms for Pi0.5. A public-internet round trip on top of that turns a working policy, or a working operator, into a hesitant one. See inference latency.

Pay and market size: what can be sourced

This is where most articles invent numbers. There is no occupational classification for robot teleoperator and no government wage series, only individual job postings, a few news reports, and collection statistics inside research papers. Here is everything in that category that survived being checked.

What you can infer is direction, not magnitude. The pi0 pre-training mixture was 10,000 hours of dexterous manipulation data from 7 robot configurations and 68 tasks, with a further 9 percent of its timesteps from Open X-Embodiment, Bridge v2 and DROID. Open X-Embodiment is itself 22 robots from 21 institutions covering 527 skills, every hour of it recorded by a person at a rig.

How to start, two ways

Less alternatives than a sequence. Recording your own dataset tells you whether you like the work; a platform is how you get paid without buying hardware or finding customers.

Is this a job with a future?

It is a job whose shape keeps changing. Current vision-language-action models are data-hungry in a way only human demonstration satisfies, and nothing on the horizon removes that for contact-rich manipulation. What changes is the mix: as policies handle the easy cases, the valuable hours shift from bulk recording toward the episodes where the robot failed, which are exactly the episodes nobody has.

For background, the RoboTurk write-up covers where crowdsourced teleoperation started and the VLA overview covers what your episodes feed. The arena holds 85 models and 332 benchmark results, each value linked to its source.