Learning and force adaptation for interactive actions

Dynamic movement primitives prove to be a useful and effective way to represent an action of a given agent. However, current DMP formulations do not take the interaction among multiple agents into the consideration. We propose a new formulation called Coordinate Change Dynamic Movement Primitive (CC-DMP), which explicitly represents the relationship between leader and follower in multi-agent collaborative tasks. The formulation of the CC-DMP is motivated by the idea of learning the relationships among interacting agents instead of learning single agent movements. We extend this idea in our previous work further to include a coupling term into the formulation, which is initially learned to adapt the CC-DMP to the current environment and context and is subsequently refined to account for dynamic changes perceived by the agent's sensor system. The incorporation of the coupling term in the formulation allows learning and adaptation of CC-DMPs for interaction tasks as well as encoding the complexity of the learned actions and their adaptation to dynamic changes in the scene, specifically caused by another agent. We demonstrate our approach in the context of learning wiping movements for dynamically changing task constraints.