Safe Human-Robot Motor Skill Learning Through Probabilistic Dynamic Movement Primitives and Control Barrier Functions

We present a novel framework that enables robots to learn motor skills from human demonstrations with guaranteed safety. Our framework comprises two components: a high-level planner and a low-level controller. The planner generates a safe trajectory region through an embodied state-action mapping (i.e., policy) computed using Dynamic Movement Primitives. The policy enables the robot to adapt to new environments, avoid obstacles, and stay within the predefined safe region. To execute the planned trajectories, we combine a control Lyapunov function controller, which tracks the mean trajectory of the distribution, and a control barrier function controller to strictly enforce constraints on the position of the robot's end effector, ensuring it remains within the safe zone. Our proposed framework is tested and validated through simulations and experiments with a Baxter robot to perform a pick and place task. These experiments demonstrate that the ProDMP is capable of generating distributions from demonstrated trajectories. Moreover, utilizing our proposed controller, the robot can navigate within these distributions from starting to goal locations while avoiding obstacles and reduces overall control effort.