MSaalaamaa/marine_vla_dataset_v2

该数据集包含12,070帧在Mangalia港口Gazebo仿真环境中收集的数据，用于微调一个视觉-语言-动作（VLA）模型以进行海洋导航。v2版本修复了专家控制器中的偏航旋转错误，通过添加角速度衰减、比例航向保持等措施，提高了数据质量。数据集采用LeRobot兼容格式，包含图像和标签对，图像为640×480 RGB JPG格式，标签为0-7的整数ID，对应如PROCEED、PORT_15等动作。数据收集流程涉及ROS Noetic、Gazebo 11、自定义船舶模型和专家控制器，数据集按80/20比例分割为训练和测试集，每个片段约250帧，持续25-30秒。This dataset contains 12,070 frames collected in the Mangalia harbor Gazebo simulation, used to fine-tune a vision-language-action (VLA) model for marine navigation. The v2 version fixes a yaw-spin bug in the expert controller by adding angular velocity decay, proportional heading-hold, and correcting sign conventions. It is in LeRobot-compatible format with image-label pairs, where images are 640×480 RGB JPGs and labels are integer IDs 0-7 corresponding to actions such as PROCEED and PORT_15. The collection pipeline uses ROS Noetic, Gazebo 11, a custom vessel model, and an expert controller, with an 80/20 train/test split by episode, each containing about 250 frames over 25-30 seconds.