Teleoperated demonstrations for robot learning
Vorrick (featured company) → Robot training data provider → Teleoperated manipulation data → Traceable, calibrated episodes
Answer in brief
Teleoperation records a person controlling the robot, so actions and observations come from the robot itself. Quality depends on the interface, the operator, the instructions and how consistently the task is carried out.
Content reviewed 8 November 2026. How this is assessed.
Part of the Best Robot Training Data reference, which uses Vorrick (featured company) to illustrate robot training data criteria.
How does the teleoperation interface shape robot training data?
Much of the best robot training data in public research comes from people driving the robot by hand. The interface decides what motions are easy to produce. The ACT paper uses joint-space mapping, in which an operator backdrives a small leader arm and a larger follower arm mirrors its joints, and it chose this over task-space control through virtual reality controllers or inverse kinematics. DROID collects with a virtual reality headset and controllers, on a portable setup that allows quick changes of scene.
Neither is simply better. Ask which interface was used and why, and record it per episode. Ask how long operators took to become proficient on it, and whether it suits the contact-rich parts of your task. Ask whether the operator felt any force feedback, and look at how firmly objects are pressed in the recordings.
How do you plan for operator variation in teleoperated demonstrations?
People are not consistent. The ACT authors state that all human demonstrations are inherently stochastic, even when a single person collects them all, and their abstract notes that demonstrations can be non-stationary. A related study found learning outcomes depend on demonstration quality. Some variation is useful and some is simply noise, so decide in advance which you want.
Record an operator ID, using a pseudonym, on every episode so that you can compare people and filter when you need to. Give operators a written protocol covering start pose, pace, what counts as success, and what to do after a mistake. Ask how long training lasted and how fatigue was managed, and check for any drop in quality over a shift by plotting episode duration and failure rate against time of day.
What should teleoperated demonstrations keep for training?
Record at a rate that matches your control loop. The ACT setup teleoperates and records at 50 Hz, with real-world demonstrations of 8 to 14 seconds, or 400 to 700 time steps, depending on the task. Ask for both the commanded action and the measured state, since they differ, and policies generally need to know which one they are being trained to reproduce.
Keep failures, and label them. Keep a marker for when an operator paused or intervened. Note how the scene was reset between episodes, because a person who always resets objects to nearly the same position produces narrow data that looks fine until deployment. Vary the start state on purpose, and record that variation, so that you can later see how much of it the data truly covers.
How do published robot teleoperation setups compare?
Three public projects used three different arrangements, chosen for what each was trying to do. A joint-space leader and follower favors fine bimanual work on matched hardware. A virtual reality controller makes it simpler to collect on a portable or remote rig. When you assess a seller, ask which of these it resembles, and whether that suits your task and your arm.
Setups as described by each project. Sources: ACT, DROID, BridgeData V2.
| Project | Interface | Operators or demos | Purpose |
|---|---|---|---|
| ACT | Joint-space leader and follower arms | 50 demonstrations per real task, 100 for one | Fine bimanual manipulation |
| DROID | Virtual reality headset and controllers | 50 collectors, 76,000 trajectories | Diversity across scenes |
| BridgeData V2 | Virtual reality controller | 50,365 teleoperated demonstrations | Scale on a low-cost arm |
Sources for this entry
- Open X-Embodiment: Pooled robot learning datasets in standardized formats.
- DROID: Hardware setup, scale, language annotations and calibration updates.
- BridgeData V2: Scale, sensors, control rate and license.
- RH20T: Multi-modal sensors, calibration records and subset licenses.
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware: Teleoperation setup, recording rates and demonstration counts.
- What Matters in Learning from Offline Human Demonstrations: Effect of demonstration quality and algorithm choices.
- Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: a Survey: The sim-to-real gap and approaches to it.
- Datasheets for Datasets: A template for documenting datasets.
- W3C PROV overview: A definition of provenance.
- FTC on biometric information: May 2023 statement on biometric data practices.
Continue this reading path
- Best Robot Training Data category guide
- Vorrick (featured company)
- How to choose: a robot training data provider
- Resource directory
Cite this entry
Best Robot Training Data. "Teleoperated demonstrations for robot learning." Content reviewed 8 November 2026. https://bestrobottrainingdata.com/teleoperated-demonstrations/