From PID to Intelligent Control: Learning-Based Controller Design for Industrial Manipulators with Sensor Fusion

Abstract

Modern production lines expect industrial manipulators to keep sub-millimetre accuracy and short cycle time while loads, friction, and ambient conditions change. This review examines practical control choices and how they interact with mechanical limits and perception. The survey covers peer-reviewed studies from 2021–2025 and two Chinese reviews with industrial relevance. The methods include structured screening of the controller series, proportional-integral-derivative baselines, adaptive and sliding mode designs, as well as reinforcement learning superposition, and evidence regarding structure-related bandwidth, stiffness, and clearance. There is also multi-sensor fusion for state and contact estimation. The analysis shows that most plants retain a stable inner loop, add feed-forward and simple observers, introduce adaptive or robust layers when uncertainty grows, and place learning above the stabilising loop where parts and scenes vary. Perception improves global alignment and contact safety but requires clear latency budgeting. The review provides scenario-level recommendations that balance accuracy, cycle time, and availability.