The judgment behind every touch.
To turn the skill on Korea’s factory floors into robot learning, we need to record the reasons behind the movement.
INTERACT ROBOTICS · Insights · Reviewed September 7, 2026
Video shows the movement. Skill holds the reason.
When a connector catches, a skilled worker doesn’t simply push harder. They shift the hand supporting the cable, retreat slightly, and adjust the angle. A small motion contains a larger decision: reading resistance and choosing what to do next.
Recording contact work means preserving the conditions behind that decision. Where contact happened. How much force was applied. When the worker stopped, tried again, and got a different result. These are records for understanding contact, beyond copying motion.
Why Korea’s factory floors?
Even highly automated countries have tasks that remain in human hands. IFR ranks Korea highest in industrial robot density per manufacturing employee. Yet high automation does not mean every process is automated. Changing tolerances, flexible parts, and work between installed pipes and wires present distinct challenges.
For us, Korea brings these problems close to the people who know how to solve them. Automotive harnesses and seals, shipbuilding hoses, precision aerospace fastening, and energy-system maintenance all depend on reading and responding to contact. Lessons from one industry may open possibilities in another.
References
- IFR · Global Robot Density in Factories Doubled in Seven Years ↗
Context on manufacturing robot adoption. The potential for recording industry-specific contact tasks is our interpretation.
Put skilled people at the center of the data.
Good demonstrations depend on people working naturally. Heavy devices, gloves that obscure touch, or delayed contact feedback can turn a record of skill into a record of adapting to hardware.
Transparency and comfort are therefore data-quality questions. We need interfaces that preserve force control, timing that connects multiple senses to one event, and context about the operator’s attention. Alongside the number of demonstrations, we ask what each one captures.
There is more to learn than success.
Record only easy insertions and you miss the judgment needed when a connector catches. Vary materials, tolerances, starting poses, and cable tension. Label recovery steps, such as retreat and realignment, as carefully as successful motion.
ContactBench maps that acquisition scope. Dream Catcher captures the demonstrations; Dream Set organizes them for Morpheus learning and field evaluation. The aim is to make previously hard-to-observe clues to human judgment learnable, one at a time.