CHINZY / FDE and robotics acceptance fieldbook / 09
Robot workcells and data capture
Freeze the workcell, fixture and capture settings using samples or replay logs. Judge motion completion, correct labels and reusable data packages separately. Retain exceptions by batch.
Acceptance checks
Original grasp and retry
Use agreed start and end actions for each grasp. A successful retry does not erase the first failure; report first-attempt success separately from final delivery.
02Matched parts and poses
Match part ID, pose and fixture revision between runs. Changing both part and gripper prevents attributing the difference to one factor.
03Placement error in a common frame
Compare target and measured positions in the same frame and unit using agreed per-axis tolerances. An unknown transform is not zero error.
04Handling success and part damage
After transfer inspect the part and contact surfaces against agreed appearance criteria. Delivery with a new scratch can pass the motion step but fail part acceptance.
05Action and frame alignment
Pair each action with the nearest image on the same timebase. Compare absolute offset with the project’s agreed window; row numbers alone do not align streams.
06Missing versus imputed frames
Find gaps in recorded sequence numbers. Tag missing and postprocessed values separately. An imputed frame may aid playback but is not a sensor observation.
07Task success versus dataset yield
A failed task can produce a complete sample. Check required fields, timing and labels; use separate numerators for data validity and task success.
08Independent dataset replay
Read a sample offline in an independent environment using the manifest, schema revision, calibration and checksum. A vendor video alone is not reproducible delivery.