checo1092/Continual_Learning_for_Autonomous_Driving
收藏资源简介:
Carla Doker Final Packed 是一个用于自动驾驶领域持续学习的基准数据集,基于CARLA 0.9.15模拟器构建。该数据集包含7个不同的城镇,每个城镇有7个紧凑域(涵盖不同天气和NPC条件),共49个基础任务,490个episodes,总计490,000帧数据。数据以tar分片格式存储,包含RGB图像、语义分割标签、LiDAR点云和车辆状态(state.csv)等多种模态。数据集采用固定的空间划分(训练、验证、测试集),确保同一驾驶走廊在不同条件下不会同时出现在训练和评估中。该数据集旨在研究领域转移下的持续学习,评估自动驾驶感知和控制模型的灾难性遗忘问题,并支持基于KAN的模块化持续学习实验和从多模态数据进行的模仿学习基线研究。
Carla Doker Final Packed is a benchmark dataset for continual learning in autonomous driving, built on the CARLA 0.9.15 simulator. It includes 7 towns, each with 7 compact domains (covering various weather and NPC conditions), resulting in 49 base tasks, 490 episodes, and 490,000 frames. The data is distributed as tar shards and contains multiple modalities: RGB images, semantic segmentation labels, LiDAR point clouds, and vehicle state (state.csv). The dataset uses a frozen spatial split (train, validation, test sets) to prevent the same driving corridor from appearing in both training and evaluation under different conditions. It is designed for continual learning under domain shift, evaluating catastrophic forgetting in autonomous driving perception and control models, and supports KAN-based modular continual learning experiments and imitation learning baselines from multimodal data.





