A 3D Simulation Framework for Behavior Cloning on Edge AI-Enabled E-Scooters in Smart Cities

This work presents a 3D simulation framework for training and evaluating Behavior Cloning (BC) models on edge AI-enabled e-scooters. The proposed approach uses binary vi-sual input and lightweight neural policies to enable autonomous navigation under constrained computational resources. Using the CARLA simulator, a dataset was collected in Town01with four discrete control actions (forward, left, right, stop), ensuring balanced maneuverrepresentation. Several visual backbones were explored, including ResNet50, MobileNetV3,and DINOv2, combined with MLP and XGBoost policy heads. Experimental results showthat the ResNet50+MLP configuration achieved the highest performance (99.6% accuracy,0.996 weighted F1) while maintaining real-time inference (140 FPS) with only 94 MB ofmodel size. These findings demonstrate that reliable visual imitation can be achieved evenwith simplified inputs and modest hardware, enabling practical deployment of autonomousor teleoperated e-scooters within smart city environments.