Robot demonstration data is useful when another person can understand what was captured and whether it fits the intended task. A large number of recordings does not, by itself, establish coverage, quality, or readiness for training. Start with a defined behavior and an acceptance checklist before increasing collection volume.
This guide is a planning framework for a robotics team. Sensor formats, calibration methods, and safety procedures depend on the equipment and research setup. Use the relevant hardware documentation and your team's review process for implementation.
Define a complete episode
Write the starting state, intended action, target outcome, and conditions that end a recording. Decide how resets and interruptions are represented. A dataset becomes difficult to use when one collector treats a reset as a new episode and another leaves several attempts in the same file without markers.
Include the objects and environment conditions that matter. For a manipulation task, record object variation, placement, camera views, and whether the final state can be judged from the captured data. For a navigation task, describe the route and environment changes.
Make the target behavior explicit enough that two reviewers can independently agree whether an episode met it. If they disagree, refine the definition before collecting thousands of additional examples.
Record the capture setup
Keep a versioned description of cameras, sensors, control interface, capture software, and relevant configuration. Store calibration records and the time they were checked. Label changes in the setup so downstream work can distinguish data captured before and after an adjustment.
Document units, coordinate frames, timestamps, sampling behavior, and missing-value conventions. These details are often more useful to a new dataset user than a polished description of the collection hardware.
Check whether streams are synchronized well enough for the intended use. Define an acceptable tolerance with the research team and test it using a known event in the capture setup. Avoid claiming a universal synchronization threshold that has not been justified for the task.
Review examples before collecting at scale
Inspect a small batch end to end. Play back recordings, check sensor continuity, verify episode markers, and confirm that the desired outcome is visible. Try the downstream reader or training-data conversion process on this batch.
Record defects separately: missing frames, mislabeled outcomes, clipped motion, wrong units, incorrect calibration, and an interrupted control stream. Decide which defects require recollection and which can be represented honestly in metadata.
Keep a review log containing the episode identifier, issue, reviewer, and resolution. Do not silently repair a label without preserving the reason. A dataset's revision history helps explain why an experiment changed after an update.
Keep success and failure examples distinguishable
Failure examples can be valuable when the intended method can use them, but an unlabeled failure should not appear as a successful demonstration. Define the outcome categories and record why an episode ended.
Capture variation deliberately rather than accidentally. A coverage matrix can list object type, starting position, environment condition, collector, and outcome. The matrix reveals missing combinations; it is not proof that the resulting model will generalize to unobserved conditions.
Keep evaluation examples separate according to the research protocol. Reusing an evaluation episode in training can make a reported result difficult to interpret. Document the split and the unit of separation, such as scene, object, or collection session.
Review permissions and handling
Collect only in approved environments with the applicable equipment and site procedures. Avoid capturing unnecessary faces, private documents, or confidential screens. Agree on access, retention, and permitted uses before sharing recordings outside the team.
An equipment purchase does not automatically grant rights to every object, environment, or person visible in the data. Record those boundaries alongside the technical metadata so a later collaborator does not have to infer them.
Ask concrete equipment questions
When discussing a collection interface with Droid, ask which data it captures, how timestamps are generated, how calibration is performed, and how the output is exported. Request a representative sample that your team can inspect through its actual pipeline.
Pair these questions with the site survey checklist when collection involves a mobile platform. Choosing equipment around a defined capture and review process is more reliable than assuming additional recording volume will solve unclear task definitions.