Convolutional Neural Networks with Stable-Baselines for Optimized Path Planning by Simulation

Robust perception and intelligent path planning are critical for reliable autonomous operation in robotics and autonomous vehicles (AVs), especially in complex environments. AVs must accurately analyze their surroundings, detect obstacles, and select safe routes, as even minor errors can cause collisions. Convolutional neural networks (CNNs) have greatly enhanced road understanding and real-time decision-making. This study integrates a CNN model into Stable Baselines 3 (SB3), enabling interaction with a SUMO simulation map to determine optimal routes using diverse metrics, including minimizing travel time, reducing congestion, avoiding traffic lights, preventing loops and dead ends, maintaining shorter distances to destination, and ensuring stable speeds. This addresses key limitations in prior research and supports more efficient and sustainable autonomous mobility. This work makes three main contributions. First, it introduces a multi-input CNN with a VGG-like backbone capable of fusing spatial traffic features with scalar speed data for path planning. Second, it identifies shortcomings in existing CNN-based approaches, including limited multi-modal fusion, insufficient handling of traffic-light congestion, and inadequate loop-avoidance mechanisms. Third, empirical evaluation in SUMO shows that multi-input fusion yields more stable predictions and superior routing compared to single-input CNN baselines, demonstrating the value of multi-modal deep learning for autonomous mobility.