Adaptive Robotic Control in Automotive Assembly: A Sensor-Fusion and Simulation-Based Framework

Abstract

Introduction: The increasing complexity of automotive manufacturing, driven by customization requirements and tight assembly tolerances, necessitates robotic systems with enhanced flexibility and real-time adaptability. The automotive industry has long been a leader in implementing industrial automation, particularly through the deployment of robotic systems to streamline repetitive manufacturing tasks. However, with the increasing complexity of vehicle models, shorter product life cycles, and growing customization demands, conventional rigid robotic systems are no longer sufficient. These systems, traditionally optimized for fixed tasks, struggle to cope with dynamic environments, part variations, and unstructured conditions inherent in modern manufacturing floors. To address these challenges, recent advancements in adaptive robotics—enabled by sensor integration, machine learning, and simulation-based control—offer promising avenues for developing highly flexible, precise, and intelligent robotic solutions [1], [2].

Objectives: This paper presents a simulation-based adaptive control framework that integrates a three-layer architecture: (i) a perception layer using vision and force sensors, (ii) a control layer incorporating an adaptive PID controller with reinforcement learning, and (iii) an execution layer involving robotic actuation and feedback.

Methods: The system was implemented in MATLAB/Simulink and tested on a 6-DOF robotic manipulator model with simulated sensor noise, external disturbances, and part misalignment. Sensor data is fused via a Kalman filter, enabling accurate estimation of pose and contact force. While Q-learning is used for controller adaptation in simulation, the architecture is designed to accommodate real-time implementation using step-wise learners such as SARSA(λ) or offline-trained gain schedulers.

Results: Results demonstrate a 76.6% reduction in RMSE and a 50% improvement in adaptation time over static PID control. The framework provides a scalable and robust approach suitable for autonomous robotic assembly systems.

Conclusions: The proposed three-layer architecture—comprising perception, control, and execution—enables the system to dynamically respond to variable part geometries, sensor noise, and external disturbances. Simulation results demonstrate that the adaptive controller, when combined with Kalman-filtered sensor data and Q-learning-based gain tuning, significantly outperforms traditional static PID controllers in both accuracy and robustness. Specifically, the system achieves a 76.6% reduction in RMSE and exhibits faster recovery under disturbance conditions.