Intelligent Walrus Optimizer Fused Feedforward Neural Network (IntWO-FFNet) for Embedded Perception and Decision-Making in Industrial Robots

Abstract

Intelligent industrial robots rely significantly on accurate vision and autonomous decision-making to do high-performance tasks. Embedded systems, compact, real-time computer units, have become critical for delivering these capabilities, especially in resource-constrained industrial environments. Despite their advantages, embedded systems meet obstacles such as high computational cost, overfitting, and inadequate parameter tuning, which impede real-time performance and generalizability in dynamic industrial environments. The purpose of this research is to develop an embedded neural network framework that has been tuned using metaheuristic algorithms to increase the precision of robotic vision and the effectiveness of decision-making while considering available resources. Multimodal data (vision, force, and proximity) are acquired from industrial environments. Raw data is cleaned and normalized using min-max scaling. Principal Component Analysis (PCA) is used to extract statistical and geographical characteristics, reducing dimensionality. This research proposes an Intelligent Walrus Optimizer Fused Feed-forward Neural Network (IntWO-FFNet) to enhance the accuracy, efficiency, and adaptability of industrial robots by enabling intelligent perception and decision-making on embedded systems. An FFNet is used on the embedded system to identify environmental inputs and forecast task-specific behaviors. The FFNet is fine-tuned with IntWO to improve learning rates, weight initialization, and hidden layer configurations for less error and faster convergence. The proposed method was implemented using Python 3.10.1. The proposed IntWO-FFNet approach performs better than multimodal baseline architectures, achieving superior results, with accuracy ranging from 95% to 99%. Integrating neural networks with optimization approaches into embedded systems dramatically improves real-time robotic perception and decision-making, providing intelligent automation aligned with industrial robots. The dataset contains 10,214 multimodal samples across five task classes (pick, place, weld, idle, interaction) and five industrial object types. All experiments were executed on a simulated embedded environment (Raspberry-Pi–equivalent ARM setup) using Python 3.10.1, with evaluation performed using stratified train–validation–test splits. Performance was benchmarked against the BIIRCS baseline model, where the proposed IntWO-FFNet achieved 96.3% accuracy, 9.8% RMSE, and 97.5% task-coverage rate.