Dynamic Movement Primitives in Constrained Environments: Teaching Control Policies Through Repeated Demonstrations

Enabling robots to learn from few demonstrations is a crucial step toward task automation. This paper addresses a key challenge in robotic learning from demonstration: developing adaptable, constraint-aware systems that learn from a limited number of demonstrations, infer task-specific geometrical constraints autonomously, and ensure trajectory adherence to these constraints while remaining accessible to non-expert users. The proposed methodology integrates automatic task-constraint extraction from teleoperated demonstrations with a novel sigmoidal coupling term (SIG-CDMP) to enforce spatial constraints within Dynamic Movement Primitives (DMPs). By analyzing variability in demonstrations, the framework defines a tolerance zone for robot motion and ensures that generated trajectories remain within these bounds, even when adapting to new initial or goal positions. The efficacy of this approach is validated in two real-world industrial applications—sterility testing in pharmaceutical industry and cable wiring in electric vehicle batteries—demonstrating negligible increases in computational cost and smooth, constraint-compliant trajectories. By integrating task-constraint extraction and enforcement, this approach advances the development of constraint-aware robotic systems that learn from repeated demonstrations to teach control policies respecting inferred geometrical constraints, paving the way for safe and reliable task automation in complex environments.