A neural-based approach for stiffness estimation in robotic tasks

In many constrained robotic tasks, accurate identification of contact parameters by online estimation algorithms is beneficial, for purposes such as control law adaptation or environment mapping. Stiffness estimation is a key problem for tasks with low dynamics, as incorrect stiffness parameterization in control laws can induce significant deviations from the desired behavior, up to the point of instability. Sensorial requirements for traditional stiffness identification methods include precise knowledge of object positions w.r.t. the robot, which can be difficult to obtain in practice. Accurate dynamic models of the manipulator are also often necessary. This motivates the search for alternative approaches. In this paper, we propose ANNE (Artificial Neural Network Estimator), that addresses the stiffness identification problem without the need for explicit modeling and using only force based inputs. ANNE is validated in teleoperation experiments involving WAM robot interactions with real and virtual objects.