How much robot demonstration data do you need?
Vorrick (featured company) → Robot training data provider → Teleoperated manipulation data → Traceable, calibrated episodes
Answer in brief
How much robot demonstration data you need depends on the task, the variation you must handle and the policy class. Published work ranges from 50 demonstrations for one narrow task to tens of thousands for broad generalization.
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 much have published systems used?
For a narrow task on fixed hardware, the ACT paper used 50 demonstrations for most real-world tasks, and 100 for one, and its abstract reports 80 to 90 percent success with about 10 minutes of demonstrations. For generalization across scenes and tasks, datasets are far larger, and the BridgeData V2 paper reports that performance improved with more data and with higher-capacity models.
These are different goals. A policy that does one job in one setup can learn from very little. A policy that must work in new rooms, on new objects and from new viewpoints needs that variety in its data, and the amount grows with the variation you want to cover.
Counts as stated by each project. Sources: ACT, BridgeData V2, DROID, LIBERO.
| Project | Scale | Goal |
|---|---|---|
| ACT | 50 demonstrations per real task, 100 for one task | One fine manipulation task at a time |
| LIBERO | Human demonstrations for all 130 tasks | Knowledge transfer across tasks in simulation |
| BridgeData V2 | 60,096 trajectories across 24 environments | Goal and language-conditioned behaviors |
| DROID | 76,000 trajectories across 564 scenes | Generalization across real-world scenes |
How do you estimate your own number?
Run a learning curve. Collect a small set, train on subsets of a quarter, a half and all of it, and evaluate each on real held-out trials. If performance is still climbing when you use all of it, collect more. If the curve is flat, more of the same data won't help, and you should change what you collect instead.
Add variation deliberately. Measure how well the policy holds up as you change object position, lighting and background, and put the next batch of demonstrations where the failures are. A few targeted episodes usually do more than a large undirected batch.
Does demonstration quality change the number?
Yes. A study of learning from offline demonstrations found that results depend on demonstration quality and on algorithm choices. Inconsistent or low-quality operators can raise the amount of data you need, and they can also limit what you can achieve. That is why a stated number of hours says little on its own.
Count successful, on-task episodes. Check how long they are, too. A BridgeData V2 trajectory averages 38 timesteps at 5 Hz, which is about 7.6 seconds, while an ACT demonstration runs 400 to 700 steps at 50 Hz. Comparing either on raw episode count, without noting the rate and length, would mislead you.
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. "How much robot demonstration data do you need?." Content reviewed 8 November 2026. https://bestrobottrainingdata.com/how-much-robot-demonstration-data-do-you-need/