<p>The increasing frequency of natural disasters and terrorist attacks highlights the critical need for efficient evacuation planning in buildings, where outdated or inaccurate floor plans often hinder emergency response efforts. To address this, we propose a new hybrid framework intervening Improved Harris Hawk Optimization (iHHO) for adaptive path finding, a T2T-Vision Transformer (T2T-ViT) model for pixel classification of walls/doors in floor plans, and Cubic Bezier Curve Smoothing (CBS) for evacuation routing free of any collision. Our method optimally balances exploration and exploitation in path planning. Whilst generating smooth trajectories free from obstacles. After rigorous testing on the CubiCasa5K dataset, this method was found far superior to CNN, DNN, and 2D-GA models, presenting results with precision = 98.8%, sensitivity = 98.8%, and specificity = 98.8%, respectively, when trained with 90% of data in 100 iterations. The framework's scalability and real-time capability give it mainstream application in smart building safety, emergency management, and robotic navigation. A step forward in the intelligent autonomous evacuation systems for dynamic environments is thus provided by this work through the integration of advanced machine learning and metaheuristic optimizations. Highlights. A technique is proposed to find the shortest evacuation routes through floor plans. The method employs he hybrid of iHHO and T2T-ViT for the generation and classification of routes. The proposed method exhibited more than average accuracy, sensitivity, and specificity on the setups tested. The method outperforms CNN, DNN, and 2D-GA in complex floor-plan navigation.</p>