reactor_x2_100
收藏资源简介:
该数据集名为reactor_x2_100,包含79个episodes,共6818帧,来自so101_follower机械臂数据。每个episode是通过将不同的干扰物-水果组合(如苹果、香蕉、葡萄、猕猴桃、柠檬、桃子、梨、李子、草莓等)添加到单个真实记录的拾取和放置episode中,并使用Reactor XMAX X2视频编辑生成。数据集任务为抓取橙子并放入盘子。数据来源于一个真实记录的抓取橙子episode(LightwheelAI/leisaac-pick-orange,episodes/orange_ep0),机器人使用SO-101 follower arm,6自由度动作/状态,单前视摄像头(480x640,30fps)。构建方法:通过脚本重新渲染源episode 95次,每次添加不同的干扰物水果和光照风格(金色(温暖午后)、阴天、正午、台灯、早晨),同时保持原始橙子、盘子和机械臂运动不变。输出帧经过每帧存活门控,仅保留目标橙子仍可识别的帧,场景被截断至目标持续不可识别的第一个连续帧之前。最终79个场景幸存,16个场景因产生帧数过少而被丢弃。数据集统计信息:episodes 79,帧数6818,FPS 30,动作/状态维度6,摄像头为observation.images.front(480x640),平均每帧覆盖率0.873,平均目标ΔE(中位数/90百分位)为20.7/24.8,10个episodes被门控截断,保留帧与渲染帧比例为6818/7108(96%)。每个episode的溯源信息(场景ID、提示、覆盖率、ΔE、截断、源帧范围)记录在meta/augmentations.jsonl中。注意事项:该数据集仅增强现有动作标签的episode,X2仅处理像素,不提供动作、本体感觉或相机姿态。在此增强数据上训练是否改善策略尚未验证。
The dataset is named reactor_x2_100, containing 79 episodes with a total of 6818 frames, sourced from the SO-101 follower robotic arm data. Each episode is generated by adding different distractor-fruit combinations (e.g., apple, banana, grape, kiwi, lemon, peach, pear, plum, strawberry) to a single real-recorded pick-and-place episode, using Reactor XMAX X2 video editing. The task is pick the orange and place it into the plate. The data originates from a real-recorded orange-picking episode (LightwheelAI/leisaac-pick-orange, episodes/orange_ep0), with the robot using a SO-101 follower arm, 6-DoF action/state, and a single front-view camera (480x640, 30fps). Construction method: The source episode is re-rendered 95 times via a script, each time adding different distractor fruits and lighting styles (golden (warm afternoon), cloudy, noon, desk lamp, morning), while keeping the original orange, plate, and arm motion unchanged. Output frames are subjected to per-frame survival gating, retaining only frames where the target orange is still recognizable, and scenes are truncated before the first consecutive frame where the target becomes unrecognizable. Ultimately, 79 scenes survive, and 16 scenes are discarded due to insufficient frames. Dataset statistics: episodes 79, frames 6818, FPS 30, action/state dimension 6, camera observation.images.front (480x640), average coverage per frame 0.873, average target ΔE (median/90th percentile) 20.7/24.8, 10 episodes are gated-truncated, retained frame to rendered frame ratio 6818/7108 (96%). Provenance information for each episode (scene ID, prompt, coverage, ΔE, truncation, source frame range) is recorded in meta/augmentations.jsonl. Note: This dataset only augments existing action-labeled episodes; X2 processes only pixels and does not provide actions, proprioception, or camera pose. Whether training on this augmented data improves policy has not been validated.
数据集概述:reactor_x2_100
基本信息
- 许可证:CC-BY-4.0
- 任务类型:机器人操作(robotics / manipulation)
- 标签:LeRobot、机器人操作、合成增强、reactor-x2
- 数据格式:Parquet(位于
data/*/*.parquet)
数据规模
- 总片段数:79 个(原尝试 95 个场景,其中 16 个因有效帧不足被丢弃)
- 总帧数:6,818 帧
- 帧率:30 FPS
- 动作/状态维度:6 维(6-DoF)
- 相机:单个前置摄像头(480x640 分辨率)
目标任务
"Grab orange and place into plate"(抓取橙子并放入盘子)——所有片段共享同一真实抓取动作,仅改变目标橙子周围的场景。
数据构建方法
- 基础片段:来源于
LightwheelAI/leisaac-pick-orange数据集的episodes/orange_ep0,使用 SO-101 跟随臂(6-DoF 动作/状态) - 增强方式:使用 Reactor XMAX X2 视频编辑工具,将原始单一片段重渲染 95 次,每次添加不同的干扰水果和光照风格
- 质量控制:逐帧验证原始橙子是否保留可见(而非仅按片段平均),场景在目标不可持续识别时截断;若有效帧过少则丢弃整个场景
- 保留比例:保留帧 6,818 / 渲染帧 7,108(96%),其中 10/79 片段被截断
场景变化
- 9 种干扰水果:苹果、香蕉、葡萄、猕猴桃、柠檬、桃子、梨、李子、草莓(含单果及多果组合)
- 5 种光照风格:金色(暖色调傍晚)、阴天、正午、灯光、早晨
数据统计指标
- 平均逐帧覆盖率:0.873
- 平均目标 ΔE(中位数 / P90):20.7 / 24.8
- 元数据记录:每个片段的场景 ID、提示词、覆盖率、ΔE、截断情况及源帧范围均记录在
meta/augmentations.jsonl中;被丢弃场景的信息见meta/dropped.json
注意事项与限制
- 该增强仅基于像素级编辑,不贡献新的动作、本体感受或相机位姿数据
- 训练效果尚未验证——未与原片段的 A/B 对比实验
- 覆盖率与 ΔE 是自动化代理指标,仅表示标签与像素的匹配程度,并非视觉质量保证



