| 融合Transformer与SAC算法的无人车视觉导航模型 |
 点此下载全文 |
| 引用本文：张益敏，刘磊.融合Transformer与SAC算法的无人车视觉导航模型[J].计算技术与自动化,2026,(1):1-8 |
| 摘要点击次数: 356 |
| 全文下载次数: 345 |
|
|
| 中文摘要:为应对真实环境中自主导航任务面临的高感知力和安全性需求，提出了一种融合Transformer与SAC算法的无人车视觉导航模型。该模型利用Vision Transformer（ViT）网络提取图像特征，作为强化学习算法的状态输入，以此来提高无人车在动态环境中的感知能力。为应对视觉传感器故障导致的安全性问题，预训练雷达策略作为决策过程中的低计算成本应急备选方案。在Gazebo中进行仿真实验，结果表明，融合Vision Transformer训练的视觉导航策略收敛速度更快且成功率更高；在传感器突发故障时，各模态策略能够协调配合，保证导航的鲁棒性；所训练的模型能在不同环境中展现出良好的适应性。 |
| 中文关键词:深度强化学习 Vision Transformer 自主导航 Gazebo仿真 |
| |
| Visual Navigation Model for Unmanned Ground Vehicles Integrating Transformer and SAC |
|
|
| Abstract:To address the high perception and safety requirements of autonomous navigation tasks in real-world environments, this paper proposes an unmanned ground vehicle (UGV) visual navigation model that integrates Transformer and SAC (Soft Actor-Critic) algorithms. This model utilizes a Vision Transformer (ViT) network to extract image features, which serve as state inputs for the reinforcement learning algorithm, thereby enhancing the UGV’s perception capabilities. To mitigate safety issues arising from visual sensor failures, a pre-trained radar policy is employed as a low computational cost emergency fallback during the decision-making process. Simulation experiments conducted in Gazebo demonstrate that the visual navigation policy trained with Vision Transformer not only converges faster and achieves higher success rates but also ensures navigation stability through effective coordination of multi-modal strategies during sensor failures. Furthermore, the trained model exhibits excellent adaptability and robustness across different environments. |
| keywords:deep reinforcement learning Vision Transformer autonomous navigation Gazebo simulation |
| 查看全文 查看/发表评论 下载pdf阅读器 |
次