piper-apple-picking-1m
收藏资源简介:
Piper Apple Picking (Isaac Sim) 数据集是 LeRobot v3 系列中的一个机器人操作数据集,专注于苹果采摘任务。该数据集在 Isaac Lab / Isaac Sim 仿真环境中采集,使用 cuRobo 运动规划器控制一个 AgileX Piper 6 自由度机械臂,执行将苹果采摘并放入桶中的操作。数据集规模庞大,包含 1,056 个完整 episodes(轨迹片段)和总计 1,024,754 帧图像数据,采集帧率为 30 FPS。每个 episode 包含 1 至 5 个苹果(经过数据重新平衡,分布偏向 4-5 个苹果的场景),并且每个 episode 的目标都是将所有苹果成功放入桶中。数据经过清理,移除了 99 个因采集错误(场景中存在第五个苹果但未被跟踪或采摘)而损坏的 episodes,并用新的包含 4-5 个苹果的场景替换。观测数据包括:来自两个相机的图像(`observation.images.ego` 手腕视角和 `observation.images.front` 前置视角,分辨率 640x360,AV1 编码),以及机器人状态向量(`observation.state`),具体包含关节1-6位置、夹爪开合百分比位置(0为闭合,100为张开)、夹爪闭合程度分数(1为闭合,0为张开)、末端执行器位置(x, y, z)和姿态(roll, pitch, yaw)。数据集的一个关键额外特征是 `apple_bbox`,为每个红色苹果提供归一化的边界框坐标(相对于前置相机),格式为 `[visible, x_min, y_min, x_max, y_max]`,并具有遮挡感知能力(当苹果被夹爪或树叶遮挡,或已被采摘时,`visible` 标记为 0)。此特征可用于训练目标检测模型(如 YOLO)或生成高亮/增强版本的数据集副本。该数据集适用于机器人模仿学习、视觉伺服、运动规划以及基于视觉的物体检测与操作等研究任务。
The Piper Apple Picking (Isaac Sim) dataset is part of the LeRobot v3 series, focusing on robot manipulation tasks for apple picking. It is collected in the Isaac Lab / Isaac Sim simulation environment using the cuRobo motion planner to control an AgileX Piper 6-DOF robotic arm, performing operations to pick apples and place them in a bucket. The dataset is large-scale, comprising 1,056 complete episodes (trajectory segments) and a total of 1,024,754 image frames, captured at a frame rate of 30 FPS. Each episode contains 1 to 5 apples (rebalanced to favor scenarios with 4-5 apples), with the goal of successfully placing all apples into the bucket. The data has been cleaned, removing 99 damaged episodes due to collection errors (e.g., a fifth apple present but not tracked or picked) and replacing them with new scenarios containing 4-5 apples. Observations include images from two cameras (`observation.images.ego` for wrist view and `observation.images.front` for front view, resolution 640x360, AV1 encoded) and a robot state vector (`observation.state`), which includes joint positions 1-6, gripper opening percentage (0 for closed, 100 for open), gripper closure score (1 for closed, 0 for open), end-effector position (x, y, z), and orientation (roll, pitch, yaw). A key additional feature is `apple_bbox`, providing normalized bounding box coordinates for each red apple (relative to the front camera) in the format `[visible, x_min, y_min, x_max, y_max]`, with occlusion awareness (where `visible` is marked as 0 when apples are occluded by the gripper or leaves, or already picked). This feature can be used to train object detection models (e.g., YOLO) or generate highlighted/augmented dataset versions. The dataset is suitable for research tasks such as robot imitation learning, visual servoing, motion planning, and vision-based object detection and manipulation.
数据集概述
- 名称: Piper Apple Picking (Isaac Sim) — 1M frames, with apple bounding boxes
- 许可证: Apache-2.0
- 任务类别: 机器人学 (Robotics)
- 标签: LeRobot, Isaac Sim, Piper, Apple-Picking
数据集规模与结构
- 总帧数: 1,024,754 帧
- 总片段数: 1,056 个片段
- 帧率: 30 FPS
- 机器人: AgileX Piper 6-DOF 机械臂,配备 cuRobo 运动规划器,在 Isaac Lab / Isaac Sim 中采集
传感器与观测空间
- 摄像头:
observation.images.ego: 腕部摄像头,分辨率 640×360,AV1 编码observation.images.front: 前方摄像头,分辨率 640×360,AV1 编码
- 状态向量 (
observation.state): 包含 12 维数据,依次为:- 关节 1–6 位置
- 夹爪位置 (
gripper.pos): 百分比表示(0 闭合,100 张开) - 夹爪闭合分数 (
gripper.tau): 1 闭合,0 张开 - 末端执行器坐标 (x, y, z)
- 末端执行器姿态 (roll, pitch, yaw)
额外特征:苹果边界框 (apple_bbox)
- 形状: (5, 5),每行对应一个红苹果
- 每行内容:
[visible, x_min, y_min, x_max, y_max],坐标基于前方摄像头图像归一化 - 遮挡感知: 当苹果被夹爪/树叶遮挡或已被采摘时,
visible=0,表示该框无效 - 用途: 可用于训练检测器(如 YOLO)或生成增强/高亮数据集副本
采集细节
- 每片段苹果数量: 1–5 个(经过再平衡,倾向于 4–5 个)
- 任务目标: 每个片段中所有苹果均被放入桶中
- 清理说明: 已移除 99 个损坏片段(场景中放置了第 5 个苹果但从未被追踪/采摘),替换为新的 4–5 个苹果场景,以重新平衡苹果数量分布
数据文件配置
- 配置名:
default - 数据文件:
data/*/*.parquet




