Multi-volume Potential Fields for Effective Learning from Demonstration

Dynamic Movement Primitives (DMPs) is a prominent method that enables single-shot learning from human demonstrated motion in a deterministically robust way. The acquired motion information can be translated into distinct situations with spatial and temporal deviations, allowing for robust repetition of the task taught. In this context, collision avoidance is a crucial aspect for autonomous reproduction within changing environments. Previous work on DMPs promoted potential fields for the avoidance of physical volumetric obstacles. In this work, the approach is extended to a multibody scene with four classes of shapes represented by superquadrics: the end-effector, physical obstacles, an imaginary workspace, and an attractive goal state. As a result, the volumetric end-effector allows both translational and rotational deviations to generate more effective avoiding manoeuvres. Furthermore, the novel workspace and attractive goal shapes are introduced to increase the robustness while deviating from the original path and accelerate the convergence to the final pose. The effectiveness of the proposed method is showcased in a series of experiments including a physical Learning from Demonstration use case, in which its value for practical applications is demonstrated.