Industrial robot manipulators operating in unstructured manufacturing environments — welding, precision assembly, pharmaceutical dispensing, and agricultural harvesting — face control challenges arising from nonlinear joint dynamics, model parameter uncertainty, payload variation, and kinematic singularity proximity that conventional PID controllers address inadequately. The Adaptive Neuro-Fuzzy Inference System (ANFIS) architecture, which combines the interpretability and knowledge-encoding capacity of fuzzy logic with the adaptive learning capability of neural networks, offers a model-free control strategy capable of learning and compensating nonlinear dynamics online without requiring explicit robot dynamic models. This paper presents an ANFIS-based joint controller for a 6-DOF KUKA KR10 industrial robot manipulator, incorporating real-time joint torque estimation from current signatures and Kalman-filtered encoder feedback. The controller is designed, trained on 8,400 trajectory samples spanning the robot's full workspace, and validated experimentally against PID, Model Predictive Control (MPC), and Fuzzy-PID baselines on five trajectory types including circular, figure-8, and pick-and-place tasks at payload variations of 0-5 kg. The ANFIS controller achieves overshoot of 4.1% versus 28.1% for PID, settling time of 1.2s versus 2.4s for PID, ISE of 0.86 versus 4.82 for PID, and end-effector position tracking error within ±0.8mm — within the IEC 62061 Class 1 precision requirement for collaborative robot operations