Model-free inverse kinematics of industrial robots using angle-aware LSTM networks

Abstract

Inverse kinematics is a fundamental problem in industrial robot control, typically addressed by analytical or numerical methods that require explicit geometric models and are sensitive to singularities. This paper proposes a temporal long short-term memory network with an angle-aware loss function for model-free inverse kinematics prediction of a six-degree-of-freedom industrial manipulator. The method combines temporal conditioning on previous joint states with a sine-cosine output head that eliminates prediction discontinuities at angular wrap-around boundaries. Experiments on 180,000 samples generated from an ABB IRB 2400 robot confirm that the proposed model outperforms a standard baseline and yields smooth, physically consistent end-effector trajectories. Validation on real robot hardware remains as future work.