Find robot
training data.
67 suppliers of egocentric video, teleoperation and robot demos. Compare formats and price per hour, from $2.40.
Only teleoperation records the robot’s actions.
Six kinds of data are on sale. Each card shows what one sample looks like.
Human video helps pretraining. To train actions on your robot you need its states and commands.
Robot teleoperation
Records robot actionsJoint states and commands, recorded while a person drives the robot.
Egocentric video
No robot actionsPeople doing tasks, filmed from head height.
handhandHand and body pose
Actions inferredPose estimated from video. Every point is an estimate.
Gripper demos
Actions derivedA handheld gripper records its path and width.
Annotation
Labels onlySteps and captions added to footage you already have.
Simulation
Actions in simGenerated scenes and trajectories.
What are you training?
Pick your case. We show the data type to buy first and what it costs.
Buy teleoperation data for your exact robot. It records the robot’s joint states and commands, so a policy learns actions directly.
See 16 suppliers →Manipulation on our robot: Buy teleoperation data for your exact robot. It records the robot’s joint states and commands, so a policy learns actions directly. $28 to $60 per hour, vendor estimate.
Parallel-gripper grasping: Buy gripper demos with a matching gripper. A handheld gripper records path and width. One supplier lists multisensor capture with grippers. $10 to $12 per accepted hour.
Dexterous hands: Buy tracked hand pose, then map it to your hand. Pose comes from video, so check the error on a sample first. Timeline plus 3D hand tracking: $23 per video-hour.
Household tasks: Start with egocentric video for pretraining. Then add robot data for the tasks you deploy. $2.40 to $40 per hour.
A world model: Buy existing egocentric video in volume. Existing footage is the cheapest hour on sale. From $2.40 per hour of existing footage.
Contact-rich tasks: Add touch or force data to your demos. Video misses how hard a gripper squeezes. Fingertip tactile and wrist force/torque record it. No firm publishes a price.
Evaluation: Use simulation scenes for repeatable tests. Check results on the real robot before you trust them. Quoted per project.
Human video becomes robot data at step 4.
If the mapping to your robot fails, the episode stays a video.
Click a step, drag along the band or use the arrows.
Record. A person does the task while the camera runs.
- Record. A person does the task while the camera runs. Output: Time-stamped video.
- Align. Sensor clocks are synced and the camera calibrated. Output: Synced streams and calibration.
- Track. Hand, body and head pose are estimated from video. Output: Pose tracks with confidence.
- Map to the robot. Human pose becomes the robot’s joint or gripper space. This needs knowledge of the robot. Output: Candidate robot actions.
- Check and label. Start, end, success and gaps are checked. Output: Labelled episodes.
- Write the format. Saved as LeRobot, RLDS or similar. Output: Dataset files and checksums.
- Train and test. Tested on held-out tasks on the real robot. Output: A policy and its test results.
Send us your data spec.
Four answers tell us who can supply it. We read your spec, shortlist suppliers and plan a sample test in your format.
Nothing goes to a supplier without your yes.

Questions
Who sells robot training data?
Firms that record new data for you, label footage, sell finished datasets or run marketplaces. Filter them by data type on the supplier list. Open the list
What does an hour cost?
Asking prices run from $2.40 per hour of existing video to $60 per hour of teleoperation. See the cost page
Can I train my robot on human video?
For pretraining, yes. To train actions on your robot you need its states and commands, or a pipeline that maps human pose to it.
What happens after I send a spec?
We shortlist suppliers whose offer fits and plan a sample test in your format. Nothing goes to a supplier without your yes.
Find your data.
Send one spec. Get a shortlist and a sample plan.