Kinesthetic Teaching for Robotic Assembly

This thesis considers parameter estimation related to kinesthetic teaching, where the user grabs the robot end-effector and drags it around to teach it atrajectory. This enables non-robotic experts to program a robot. During kinestheticteaching, the robot is in contact with the human, and potentially in contact withthe environment. These interactions can be considered as uncertainties, and theycan be modelled to be described by parameters.

During kinesthetic teaching, the parameters of the human dynamics willchange with time. I present how to estimate time-varying parameters using onlineparameter estimation methods. First, I show that both the standard gradientbased estimator and dynamic regressor extension and mixing have parameterestimation error related to the derivative of parameter. Second, I present dynamicregressor extension and mixing with local polynomial approximation to improvethe estimation error.

During parameter estimation experiments it is desirable to move slowly andsmoothly to have the best performance. Though, this means that there is littleexcitation in the signal, and the parameters can be difficult to estimate. I describeregularised dynamic regressor extension and mixing which can estimate part ofthe parameter vector despite missing excitation. That method is applied to estimatethe parameters of a simple human model in a contributed paper.

Some parameters such as the surface normal and the stiffness of the human armhave geometric constraints, such as unity length or symmetric positive definiteness.I present Riemannian manifolds and their usage in parameter estimation. ARiemannian manifold-based adaptive controller was used in a contribution toincrease tracking performance.

Safety is a concern in kinesthetic teaching to protect the human as well asthe environment. I present robust control barrier functions to keep a system withparametric uncertainties safe. Though, robust control barrier functions have a lotof conservatism in its behaviour when the system parameters are time-varying. Tothat end, I present a robust control barrier function that used explicit estimates ofthe parameter to reduce the conservatism.