Dataset containing 36000 trajectories generated by an Optimal Control Problem (OCP) for a mobile manipulator with 10 degrees of freedom. The OCP is solved for an initial joint state configuration and a unique randomly chosen desired tool-center-point (TCP) target pose which lies within a certain goal region. The desired TCP position is computed by drawing samples from a uniform distribution for each coordinate. In order to obtain uniformly distributed desired TCP orientations, unit quaternions are used. Only the desired rotation matrices leading to trajectories which are not ending in singularities are kept.

A fully connected feedforward neural network (NN) is used as regression model to predict the optimal trajectories given by the OCP. The dataset is used to train, test and validate the NN. All data is normalized between [-1;1]. The input consists of the start joint position, the corresponding start TCP pose and the TCP desired target pose. In order to reduce the dimensionality of the input layer, the orientation is described via quaternions. The output consists of the optimal joint positions and velocities for each collocation point k ∈ [0; N], the joint accelerations for k ∈ [0; N−1] and the optimal end time. Here N=5 is used.