No abstract

No abstract

No abstract

Solving inverse kinematics (IK) has been an important problem in the field of robotics. In recent years, the solutions based on neural networks (NNs) are popular to handle the non-linearity of IK. However, the complexity of IK grows rapidly as the degree-of-freedom (DOF) of the robot arms increases. To address this problem, we exploit the dependencies among the joints of the robot arms, based on the observation that the movements of certain joints of a robot arm will affect the movements of other joints. We investigate the idea under a data-driven setting, i.e., the NN models are trained based on supervised learning through a given trajectory dataset. Several NN architectures are examined to exploit the joint dependencies of robot arms. A greedy algorithm is then presented to find a proper sequence of applying the joints to decrease the distance error. The experimental results on a 7-DOF robot arm show that the NN models using joint dependency can achieve the same accuracy as the single-MLP model but use fewer parameters.