Check a demonstration before you train on it
A useful episode connects the task, observations, actions and outcome. A beautiful video supplies only part of that record.
Identify data gaps and evaluation leakage before spending time on training.
Inspect one whole episode first
Choose one authorized recording and identify its start, end, task and hardware configuration. Watch it at normal speed alongside the available signals. Distinguish a complete task attempt from a trimmed success clip. Record any setup, reset or operator intervention that occurred outside the saved sequence.
Explain every field
Create a small dictionary for images, joint states, action targets, touch signals and timestamps. State units, shapes, coordinate frames and which fields are absent. A commanded position is not a measured contact force. Do not silently fill a missing tactile channel with zeros and then describe it as a real measurement.
Check time and alignment
Look for non-monotonic timestamps, repeated images, missing samples and streams with different rates. Check how observations and actions are aligned and whether a timestamp represents acquisition or arrival. A grasp with delayed touch data can teach a different relationship from the one you intended. Specify the resampling or synchronization rule rather than assuming equal row indices mean simultaneous events.
Protect the evaluation split
Decide what generalization you want to measure. Holding out episodes tests a different question from holding out objects, sessions or operators. Keep neighboring frames of the same demonstration together when creating your split. Fit preprocessing statistics on training data and document any exclusions so the evaluation does not quietly become easier.
Save a quality report, not just a dataset size
Count usable and rejected episodes and explain the reasons. Keep outcome and intervention labels when available; successful-only training data may be intentional, but it cannot establish the collection’s task success rate. With LeRobot, inspect the schema and episode metadata for the exact format version: v3 can store several episodes in one file. Confirm reuse permissions before publishing recordings.
Worked example: a five-line episode audit
| Finding | Why it matters | Next check |
|---|---|---|
| Video repeats a frame | Visual changes may appear delayed | Inspect capture timestamps |
| Action unit unspecified | Targets may be misinterpreted | Ask for the action definition |
| Same session in train and test | The intended generalization may not be tested | Review split grouping |
| Reset omitted from the clip | Assistance is hidden from evaluation | Check the full trial log |
| Touch channel absent | Touch-conditioned training is not supported | Change the task or collect that signal |
Try it yourself
Your practical task
- Open one permitted episode and identify the task, boundary and outcome.
- Write a field dictionary and mark missing observations.
- Inspect timestamps and synchronization for one motion or contact event.
- Propose a split and record what kind of generalization it evaluates.
What to produce: An episode quality report with one accepted item, one issue and one proposed split.
Check your understanding
Original sources & next steps
LeRobotDataset v3: schema, timestamps and episode metadata ↗LeRobot: imitation learning workflow ↗Original Dexhands teaching material. Lesson and linked references reviewed 2026-09-24. Research links are not endorsements or evidence of hardware compatibility.