History-Conditioned Neural Inverse Kinematics for Redundant Manipulators

Redundant manipulators exhibit an inherently non-unique inverse kinematics (IK) relationship, where a single end-effector pose corresponds to a continuum of feasible joint configurations. This set-valued structure complicates direct pose-to-joint regression and may induce discontinuities or branch switching during trajectory execution, while conventional redundancy-resolution methods often depend on iterative, model-based computations that are costly at high control rates. This paper presents a history-conditioned neural IK formulation that exploits temporal continuity to resolve redundancy. The proposed model augments the desired six-dimensional (6D) end-effector pose with the previous joint configuration, thereby transforming the globally ambiguous inverse correspondence into a locally single-valued mapping along continuous motions. A lightweight feedforward multilayer perceptron is trained to predict the current joint state from this augmented input, enabling constant-time inference suitable for real-time deployment. The approach is validated on an 8-degree-of-freedom (DOF) redundant platform comprising a Franka Emika Panda arm mounted on a prismatic linear axis. Experiments on multi-trajectory real-robot data demonstrate accurate joint prediction and stable autoregressive rollout on held-out motions, producing smooth joint evolution and consistent task-space tracking.