Our framework LOCI is a probabilistic machine learning method for interactive robot skill modulation using a task-parameterized, kernelized approach (TP-KMP): the skill is stored relative to the objects involved in the task, so it still works when those objects move.

LOCI overview: skills are learned from a few demonstrations, refined with interactive local via-points, extended to new objects through uncertainty reduction, and stretched into regions without demonstrations

A skill is first learned from a few kinesthetic demonstrations — a person physically guides the robot’s arm through the motion — four for the bearing ring-loading task in the paper. Users then correct it locally, triggered by external force, by distance, or by a button press, and those via-points are mapped into the global model so that corrections move with the object they belong to. When a skill is extended beyond its demonstrated duration, stiffness is regulated by the model’s epistemic uncertainty — its own estimate of what it has not seen demonstrated — keeping the interaction compliant where no demonstrations exist.

The approach also lets a skill be extended online: adding a camera or a new object to an existing task requires only minor local refinements rather than a fresh set of demonstrations.