Industrial robots are precise but usually need detailed programming for each motion. Teaching a robot to pack many item/container combinations by writing a separate policy for every combination quickly becomes impractical.
Example: 5 item types × 5 containers can create a large space of combinations, each potentially needing its own policy.
One-shot imitation learning means the robot learns a new task from a single demonstration. Instead of thousands of trials or hand-coded motion scripts, a human shows the task once and the system generalizes.
Complex jobs can be decomposed into reusable skills:
If those skills compose, the robot can remix them for new object/container pairs without learning every combination from scratch.
Compositional one-shot learning makes robots more adaptable in warehouses, homes, and other open-ended settings where object variety is high and exhaustive programming is unrealistic.
Takeaway: Reusable subskills plus a single demonstration can unlock many new task combinations.