Movement Primitives with Explicit Constraints for Imitation Learning of Pick-and-Place Action

Movement primitives are the basic units of motions for complex behaviors. They are widely used in imitation learning because they provide leaning and generalizing operator for demonstrated behaviors. As a trajectory-level encoding method, movement primitives have encountered problems when learning and generalizing under certain constraints. This paper proposes Movement Primitives with Explicit Constraints (MPs-EC) for the imitation learning of pick-and-place actions. This methodology learns constraints from demonstration and then reproduces them in generalized trajectories. Meanwhile, an optimization method converts the learned constraints into the scalar weight of a target object to realize constraint reproduction. In this way, the stable constraint reproduction in the generalized trajectory ensures successful tasking. A simulation of fixed orientation pick-and-place action and an experiment of dual-arm pick-and-throw action verify the effectiveness of the proposed movement primitives.