Mobile manipulation is a topic that attracts many researchers, as robotics is trying to integrate navigation activity with the capability to effectively work with objects in the environment.
This paper considers a manipulator carried on a mobile nonholonomic base, which is a quite common solution in industrial situations, where those kinds of commercial systems start being produced. What about asking the manipulator hand to follow a given trajectory in space? The solution may be found through specific heuristics, as decoupling the system in the mobile base and the arm, or reasoning on the task, or using some learning method. Adaptive motion/force control strategies are the choice in this paper.
The authors’ basic assumption is that the dynamics of the system are not exactly known a priori, so an adaptive controller that integrates a neural network to learn is a good candidate, as indicated by some reported successful examples. The proposed neural net is RFWNN, a combination of fuzzy, recurrent, and wavelet nets, with four layers, used in the controller to approximate the unknown parameters.
Section 2 starts with the definition of the dynamic equations of an n-degrees of freedom (dof) arm mounted on a mobile platform, and continues with the basic definition of the RFWNN net.
Section 3 describes the goal of the adaptive control algorithm, which is to move the joints to track the desired positions and to track the desired constraint force, and develops the algorithm. The control of force is decoupled from the control of position. The role of the net is to identify online the unknown parameters of the robot control system and to compensate for disturbances. The stability of the controller is also demonstrated.
Section 4 reports on experiments in simulation on a simple 2-dof arm and on a real platform used in the laboratory. Many pages of graphs show that the RFWNN net is better than other nets in reducing the error of the controlled system.
The bulk of the paper is about control theory. It is a good demonstration that controllers integrating neural nets are an effective solution. The neural net used is a state-of-the-art, yet quite advanced, solution. The authors conclude: “The proposed RFWNN’s control system can be applied as a good alternative in the existing mobile manipulator control system.” Unfortunately, it is not easy to replicate the experiment due to a lack of details about the network; also, it is not obvious how to apply the method to a manipulator with more dof, where the orientation of the hand is important.