Teleoperation vs. human video: which robot training data to buy in 2026
Teleoperation gives data in the robot’s own body but is slow and needs a robot per operator. Human video from wearables is several times faster to collect and scales with real work. Most programs now use both.
· 3 min read · neocognix
Key takeaways
- Teleoperation records the robot’s own joints and grippers, so it needs no retargeting, but every hour needs a robot and a trained operator.
- Wearable capture is faster: UMI collected demonstrations more than three times faster than teleoperation, and DexWild recorded 201 demonstrations an hour against 43.
- Human video transfers best on top of diverse robot pretraining, so the practical answer is a mix: human video for breadth, teleoperation for the robot’s own contact-rich skills.
- Buy human video for coverage of real tasks and real environments; buy teleoperation for fine-tuning on the exact embodiment.
The short answer
Use human video to cover many tasks, environments and operators at low cost per hour, and use teleoperation to fine-tune on your robot’s body for the skills that depend on force and contact. The two are complements, and the published results on transfer assume both.1
How the two methods compare
| Teleoperation | Human video (wearables) | |
|---|---|---|
| Embodiment | The robot’s own joints and grippers | Human hands; retargeted to the robot |
| Collection speed | Baseline | 3× faster (UMI), 201 vs 43 demos per hour (DexWild) |
| Hardware per operator | A robot plus a teleoperation rig | Cameras and a wearable rig |
| Where it can be recorded | Where the robot is | Wherever the work already happens |
| Strength | Contact-rich, embodiment-specific skills | Breadth: tasks, environments, operators |
| Main gap | Cost and speed at scale | No force data; needs a pretrained robot base |
Speed and cost
The clearest numbers come from academic rigs. The Universal Manipulation Interface (UMI) collected demonstrations more than three times faster than teleoperation on its cup-arrangement task.2 DexWild recorded 201 demonstrations an hour with a wearable setup, against 43 an hour with teleoperation.3
Teleoperation also carries labor cost on top of the robot. Tesla has advertised $25.25 to $48 an hour for operators who wear motion-capture suits to drive its Optimus robots.4 Recording skilled people who already do the job turns an existing shift into data, without building a fleet first.
What transfers, and what does not
Physical Intelligence found that adding human video with head and wrist cameras nearly doubled a pretrained model’s performance on generalization tasks that appeared only in the human data. The gain appeared only after diverse robot pretraining, and the human data carried 3D hand tracks and dense language labels.1
- Transfers well: where to reach, the order of steps, how objects are approached and grasped, what the eyes attend to before the hand moves.
- Transfers poorly: grip force, fingertip slip, and anything that depends on the robot’s own dynamics.
How labs are combining them
The spend is real: Bessemer estimates robot data costs across the industry will exceed $3 billion over the next two years.5 Figure has described Project Go-Big, a push to learn from large amounts of human first-person video.6 The common pattern is a pyramid: a wide base of human video for coverage, a middle layer of robot data across many tasks, and a narrow top of teleoperated fine-tuning on the exact embodiment and task.
- Pretrain on diverse robot data.
- Add labeled human video for the tasks and environments you want to cover.
- Fine-tune with teleoperation on your robot for the contact-rich steps.
- Measure each layer in your own evals before scaling it.
Whatever the mix, the human video only helps if it is labeled well. Our buyer’s checklist lists the measurements to ask for, and what egocentric data is covers the streams a training-grade episode contains.
Questions
- Is human video cheaper than teleoperation?
- Per demonstration, usually yes. Published wearable rigs collected data three to almost five times faster than teleoperation, and they need no robot per operator. The saving only counts if the labels are good enough for your policy to use.
- Can human video replace teleoperation?
- Not for most programs today. Human video adds breadth on top of robot pretraining; teleoperation is still how a policy learns contact-rich skills in the robot’s own body.
- What is retargeting?
- Mapping human hand motion to a robot’s end effector: wrist pose becomes gripper pose and finger spread becomes grip aperture, expressed in the robot’s base frame.
Sources
- 1Emergence of human to robot transfer in vision-language-action models · Physical Intelligence, Dec 2025
- 2Universal Manipulation Interface · Chi et al., Feb 2024
- 3DexWild: dexterous human interactions for in-the-wild robot policies · Tao et al., May 2025
- 4Tesla is hiring data collection operators for Optimus · Fortune, Aug 2024
- 5Bessemer predicts: robotics and physical AI · Bessemer Venture Partners, Apr 2026
- 6Project Go-Big · Figure, Sep 2025