Trajectory Tracking Solution of a Robotic Arm Based on Optimized ANN

Inverse kinematic solver based on machine learning sequential procedure for robotic applications

“…The sequential method demonstrates an improvement compared to the global from 4.22mm2 to 0.13mm2. Taking into consideration the joint errors, the global structure shows an MSE lower than 0.000657rad2, aligned with [9,10,11]; whereas for the sequential model the error was decreased by 51%, up to 0.000323rad2, for the testing dataset.…”

“…The network performance is calculated using the MSE. Similar to [11], the network hyperparameters are determined based on the GA, a search optimization algorithm that mimics the natural selection evolution. The optimization is performed on the number of neurons and the activation functions for each hidden layer; the score is calculated on both the training and validation performance.…”

“…A similar application was described in Habibkhan et al [10], where a shallow network with 12 hidden neurons obtained a validation MSE of 0.00038523 in joint space. The same planar robot can be found in [11], where the authors applied an ANN with 3 hidden layers {15 (tansig), 14 (tansig), 28 (logsic)} on 2000 points to obtain the robot IK. The achieved validation MSE was close to 0.00013 and to 0.0003 for the testing dataset, in joint space.…”

“…Moreover, taking into consideration that there is no standard procedure in literature to determine the ANN structure in advance and, hence, personal expertise is required which results in the high diversity of networks found in literature for similar applications. In this paper the Genetic Algorithm (GA) optimization was applied to obtain the ANN structure, similarly to [11]. The main contribution is the ML-procedure appraisal study for planar and spatial trajectories, to define a robust practice in datadriven applications for both ML and Hybrid methods.…”