Journal of Research and Development
Open Access

Abstract

A Comparative Analysis of CARLA and AirSim Simulators: Investigating Implementation Challenges in Autonomous Driving

Manav Khambhayata*

The advancement of autonomous driving technologies relies heavily on effective training methodologies for self-driving car AI. Reinforcement learning has emerged as a promising approach in this domain. In this paper, we present a comparative analysis of training strategies for self-driving cars using two popular simulators: CARLA and AirSim. We focus solely on the comparison between the two simulators by implementing them using current technology, analyzing their ease of implementation, and identifying the associated challenges. CARLA offers ease of setup and a realistic environment, while AirSim provides excellent overall performance despite its challenging setup process. However, integrating CARLA with TensorFlow poses certain difficulties. To conduct the comparative analysis, we implemented reinforcement learning algorithms on both simulators and evaluated their performance metrics, including training time, learning efficiency, and generalization to real-world scenarios. Our findings indicate that CARLA, despite its ease of setup, encountered challenges when integrating with TensorFlow due to compatibility issues. However, once resolved, CARLA demonstrated promising results in terms of learning efficiency and generalization to real- world scenarios, outperforming conventional methods. On the other hand, AirSim showcased superior overall performance but required substantial effort in setting up the simulator and configuring the environment. We provide insights into the strengths and weaknesses of each simulator and offer recommendations for choosing the most suitable training platform based on specific research requirements.

Published Date: 2024-12-23; Received Date: 2023-07-20