AdaKineNet: Adaptive Kinematic Neural Network for inverse kinematics of redundant mobile manipulators

Redundant mobile manipulators offer superior operational flexibility for logistics and warehousing applications, however, their inverse kinematics (IK) solutions remain challenged by singularity susceptibility, computational inefficiency, and physical infeasibility in conventional approaches. This paper introduces the Adaptive Kinematic Neural Network (AdaKineNet)—an adaptive physics-informed neural network (PINN) that integrates rigid-body transformations via automatic differentiation to achieve exact Jacobian evaluation. A dynamically weighted loss function balances positional accuracy, orientation precision, and joint constraints, while a modular architecture supports scalable deployment of manipulators with varying degrees of freedom through a unified parametric framework. Experimental validation on a 10-DoF mobile manipulator demonstrates significant improvements: positional and orientation errors are substantially reduced compared to numerical IK solvers and data-driven baselines, and joint-limit violations remain near negligible levels. The framework achieves real-time inference speeds, outperforming iterative methods by orders of magnitude, with accelerated convergence via adaptive gradient optimization. These advances establish AdaKineNet as an industrially viable IK solver, ensuring both physical feasibility and computational efficiency for safety-critical complex robotic control tasks, such as container unloading and agile manufacturing.