A generic task model and control strategy to support learning, robust control, and generalization of contact-rich manipulation tasks

Learning and controlling contact-rich tasks are based on a task model that is built from motion and wrench data collected from demonstration. In previous work, the procedure for choosing a proper reference frame to express the learned signals is often implicit, requires expert insight, or starts from proposed frame candidates. Furthermore, the exact location of the geometric features in the world or tool that determine this frame may be uncertain during task demonstration or execution. This article presents an automatic method to derive the optimal reference frame for the task model directly from recorded motion and wrench data. This frame maximizes decoupling between the learned signals, enabling the design of a decoupled task controller. The method is founded on screw theory, entirely probabilistic and requires no hyperparameters. The method is generic as it works regardless of whether the task involves translation, rotation, force, or moment, or any combination thereof. While other control approaches could be used, we propose a generic, decoupled, constraint-based and velocity-resolved controller to work in tandem with the derived task model. It involves a small number of task-dependent control parameters that need to be tuned. The derivation of the task model, the robustness of the controller, and the task generalization capabilities of their combination were validated for various tasks involving contact between rigid objects, including surface following and manipulation of articulated objects. The results show a good agreement between derived and assumed expert task frames as well as the effectiveness and versatility of the overall approach.