Why Data Quality Trumps Data Quantity
A dataset of 1,000 high-quality, diverse demonstrations often outperforms 10,000 noisy ones. Quality issues — sensor desynchronization, action discontinuities, task failures mixed with successes, and inconsistent labeling — propagate into trained policies and cause mysterious deployment failures.
The 15-Point Quality Checklist
Apply these checks to every batch of collected episodes before adding them to your training set.
- Sensor timestamps synchronized within 10ms
- No dropped camera frames (check frame count vs duration)
- Action values within physical joint limits
- No action discontinuities >2σ from mean step size
- Success/failure labels verified by second reviewer
- Task completed within expected duration bounds
- No operator self-corrections in final dataset
- Camera not occluded during critical phases
- Proprioception matches commanded actions
- Gripper state transitions at correct moments
- No duplicate episodes
- Metadata fields complete (task ID, operator, date)
- Episode stored in target format (RLDS/LeRobot)
- Batch statistics within expected distributions
- Random sample spot-checked visually