Training a neural network to observe the assembly mode of parallel robots
Adrián Peidró, Antonio Manuel García-Pérez, Paula Mollá-Santamaría, Marc Fabregat-Jaén, Óscar Reinoso
2023 the 8th International Conference on Mechanical Engineering and Robotics Research  (Krakow (Poland), December 8-10 2023)
Ed. IEEE  ISBN:979-8-3503-3051-9  DOI:10.1109/ICMERR59784.2023.10379872


Parallel robots have different solutions to their forward kinematic problem, i.e., their gripper can adopt different possible positions that are compatible with the displacements of the actuated joints. It is necessary to know the position of the gripper to control parallel robots in the operational space. However, usually, only these joint displacements are measured, which results in the indetermination of which of these possible positions is the one actually adopted by the gripper. Previous methods require redundant sensors to measure directly the position of the gripper or the displacements of passive joints. In this paper we propose to estimate the position of the gripper by solving the observability problem of parallel robots regarded as nonlinear dynamical systems. By taking the position coordinates of the gripper as states of the system, it is possible to reconstruct these states from the signals of inputs (control forces) and outputs (displacements of actuated joints) of the system. To that end, we propose a feedforward neural network to perform as the state observer, and we train it using thousands of trajectories. The proposed neural observer is tested on a simulated 2RPR parallel robot, and it is found that, combined with a refinement post-processing based on Newton-Raphson’s iterative method to solve the forward kinematics, the proposed method is able to exactly reconstruct the position of the gripper if the robot starts to move far from Type 2 singularities, at which different solutions merge.