ETHrobotlearning/tv-task2-clean-aug1_illumination-fixed

该数据集是LeRobot v3数据集，专门用于香蕉拾放任务，采用SO-101机器人手臂。它旨在支持视觉-语言-动作（VLA / SmolVLA）模型的训练。数据集包含1,184个episodes和139,368帧，帧率为10 FPS。任务提示包括33种不同的指令，涵盖四种参考风格：颜色直接指令、颜色否定指令、序数指令（从机器人视角）和相对位置指令（从机器人视角）。数据集基于原始数据集ETHrobotlearning/tv-task2-clean-aug1构建，并通过光照增强（亮度在0.7到1.2之间均匀分布，种子42）将episodes数量翻倍。特征包括动作（6维关节位置）、观察状态（6维关节位置）、前摄像头图像（480×640×3，AV1编码）、时间戳、帧索引、episode索引、索引和任务索引。数据集格式为LeRobot v3.0，适用于机器人操作和VLA模型训练。This dataset is a LeRobot v3 dataset for banana pick-and-place tasks using a SO-101 robot arm, built for Vision-Language-Action (VLA / SmolVLA) training. It contains 1,184 episodes and 139,368 frames at 10 FPS. Task prompts include 33 instructions covering four referring styles: color (direct), color (negation), ordinal (robot perspective), and relative position (robot perspective). The dataset is constructed from the base dataset ETHrobotlearning/tv-task2-clean-aug1 with each episode paired with an illumination-augmented copy (brightness ~ Uniform(0.7, 1.2), seed 42), doubling the episodes to 1,184. Features include action (6-DoF joint positions), observation.state (6-DoF joint positions), observation.images.front (video in AV1, 480×640×3), timestamp, frame_index, episode_index, index, and task_index. The format is LeRobot v3.0, suitable for robotic manipulation and VLA model training.