AI Robotics ·
How much of a factory shift should robots learn from?
Germany’s FMC³ Robotics introduced its Doceo system at November’s electronica trade show in Munich. Its pipeline captures human demonstrations through camera headsets, external cameras and optional touch gloves, then synchronizes the recordings for training and integration with robots. [1][5]
Problem definition and research question
When a battery electrode sheet curls or a tote arrives in an unfamiliar size, a skilled operator adapts their grip and posture. Factory recording can capture those departures from routine, but a camera may miss a contact point under a finger or the reason an operator stopped to inspect a defect. Should manufacturers record more work on the line or pay for controlled trials where people drive the robot directly?
Evidence and causal analysis
FMC³ is testing tote movement and placement at a consumer electronics plant and electrode-sheet handling at a battery factory. Both are proofs of concept. Published material supplies no comparative throughput or defect rates that establish an advantage over the existing control process. Its agreement to develop shared data infrastructure with Vector and partners also leaves portability of training records and validation history to be established in practice. In Korea, LG Electronics used human demonstrations, teleoperation and simulation to validate work on a washing-machine production task involving hanging a four-kilogram top panel. The combination supports the comparison of approaches, while the individual effect of demonstrations remains unidentified. [1][2][3][4]
| Dimension | International evidence | Korean evidence and application |
|---|---|---|
| Capture | Multiple camera views, motion and touch data [1] | Human demonstrations, teleoperation and simulation [3][4] |
| Robot task | Proofs of concept for tote placement and electrode-sheet alignment [1] | Validation on hanging washing-machine panels [3] |
| Remaining evidence | Continuous operation with failures and recovery | Shift-level repeatability after changes to parts or fixtures |
Comparison of alternatives
Plant engineering could record only approved segments during shifts, preserving regular moves and recoveries. Worker representatives should help decide what is recorded, when files are deleted and who can use them for training. Procurement should require the integrator to export raw records and labels. Budget for annotation and de-identification. Alternatively, the integrator could use an offline setup and teleoperation to repeat rare part variations and failures. That gives engineers control of contact conditions but needs fixtures and specialist operator time, and lighting or material flow in the test area may differ from the factory.
Implementation and evaluation
For one process, split training and evaluation by shift and material lot, then give both collection methods the same engineering budget. Quality tracks damage and alignment failures per thousand cycles, 95th-percentile cycle time and worker recoveries. A safety boundary breach stops the trial. Pause expansion if the model trained with factory recordings cannot match current quality across a full shift or raises cost per accepted part. Set permissions to review raw footage after an error and export training records before collecting anything.
Conclusion and open questions
A useful demonstration keeps its value on the next shift. Who is responsible for collecting replacement examples, and who pays when a part or fixture changes?