ETHrobotlearning/tv-task2-clean-aug1_illumination

这是一个用于香蕉拾放任务的LeRobot v3数据集，基于SO-101机器人手臂构建，专为视觉-语言-动作（VLA/SmolVLA）训练设计。数据集包含1,184个episodes和139,368帧，帧率为10 FPS，覆盖33个任务提示。机器人使用SO-101（so_follower），摄像头为前置摄像头（分辨率480×640，AV1编码），动作空间为6自由度关节位置，格式为LeRobot v3.0。数据集基于ETHrobotlearning/tv-task2-clean-aug1（592个episodes，已去除空闲边界帧），每个episode都配有一个亮度增强副本（亮度均匀分布为0.7至1.2，种子42），从而使数据集翻倍至1,184个episodes。任务提示包括四种参考风格：颜色（直接）、颜色（否定）、序数（机器人视角）和相对位置（机器人视角）。特征包括动作、观察状态、图像、时间戳、帧索引、episode索引、索引和任务索引。This is the LeRobot v3 dataset tailored for the banana pick-and-place task, constructed using the SO-101 robotic arm, and specifically designed for vision-language-action (VLA/SmolVLA) training. The dataset comprises 1,184 episodes and 139,368 frames at a frame rate of 10 FPS, spanning 33 task prompts. The robotic platform uses the SO-101 (so_follower), equipped with a front-facing camera (resolution 480×640, AV1 encoded), and the action space consists of 6-degree-of-freedom joint positions, formatted in accordance with LeRobot v3.0. The dataset is derived from ETHrobotlearning/tv-task2-clean-aug1 (592 episodes, with idle boundary frames removed). Each original episode is paired with a brightness-augmented copy (brightness uniformly distributed between 0.7 and 1.2, seed 42), doubling the total number of episodes to 1,184. The task prompts include four reference styles: direct color, negated color, ordinal (from the robot's perspective), and relative position (from the robot's perspective). The features included in the dataset are actions, observation states, images, timestamps, frame indices, episode indices, global indices, and task indices.