so101_eval3_all
收藏资源简介:
so101_eval3_all(又名SO-101 SmolVLA Eval 3)是LeRobot v3系列中的一个机器人学习数据集,专为ETH Robot Learning FS26项目的第一次评估(Eval 3)而创建。该数据集的核心任务是评估视觉语言动作模型(SmolVLA)在‘图像作为提示’场景下的性能,具体任务定义为:一个名为SO-101的机器人机械臂需要拾取一个空的330毫升可乐罐,并将其放置到工作区上呈半圆形排列的三张名人肖像中的一张上,目标肖像由一张参考照片和一个名称文本提示共同指定。数据集包含4,195个完整的操作序列(episodes),其中基于178个基础遥操作演示(涉及三种独立同分布名人:Yann LeCun、Barack Obama、Taylor Swift和六种布局排列),通过复杂的数据增强流程生成了25个增强变体,共计4,017个变体,覆盖了192位独特的名人,并确保了目标分配的平衡与均匀。数据模式(Schema)包含多模态观测和动作:观测部分包括来自手腕相机的RGB视频流(分辨率480x640,30 fps,H.264编码),以及作为恒定帧视频流(分辨率480x480)提供的目标名人参考照片(实现‘图像作为提示’);同时包含一个6维浮点向量表示机器人的关节位置状态。动作部分是一个6维浮点向量表示目标关节位置。数据集的提示(prompt)经过精心设计,包含三种类型:75%为‘默认’提示(明确提及名人姓名),15%为‘仅参考’提示(仅指示参考图像),10%为‘反事实’提示(文本提示中的姓名与可见肖像故意不匹配),用于引导模型学习遵循参考照片而非文本名称。整个数据集共包含933个独特的任务指令字符串。数据增强流程综合运用了多种先进计算机视觉模型和技术:使用GroundingDINO检测肖像,SAM 2.1进行掩码跨帧传播,结合面部感知的角点重排和Lanczos扭曲技术将原始肖像区域替换为来自Wikimedia和Bing爬取的192位名人照片库中的图像,并通过Reinhard颜色迁移和Alpha羽化混合保持场景光照一致性,最后使用InsightFace ArcFace进行身份验证(余弦相似度≥0.4)以确保增强质量。
so101_eval3_all (also known as SO-101 SmolVLA Eval 3) is a robotic learning dataset from the LeRobot v3 series, specifically created for the first evaluation (Eval 3) of the ETH Robot Learning FS26 project. The core task of this dataset is to evaluate the performance of Vision-Language Action Model (SmolVLA) in the "image-as-prompt" scenario. The specific task is defined as follows: a robotic arm named SO-101 needs to pick up an empty 330ml Coca-Cola can and place it onto one of three celebrity portraits arranged in a semicircle on the workspace, where the target portrait is specified jointly by a reference photograph and a textual name prompt. The dataset contains 4,195 complete manipulation episodes. Based on 178 baseline teleoperation demonstrations (covering three independent and identically distributed (i.i.d.) celebrities: Yann LeCun, Barack Obama, Taylor Swift, and six layout arrangements), 25 augmented variants were generated via a sophisticated data augmentation pipeline, resulting in a total of 4,017 variants spanning 192 unique celebrities, with balanced and uniform target assignment ensured. The data schema includes multimodal observations and actions: the observation part includes an RGB video stream from the wrist-mounted camera (resolution 480x640, 30 fps, H.264 encoded), and a fixed-frame video stream (resolution 480x480) providing the reference photograph of the target celebrity to implement the "image-as-prompt" setup, as well as a 6-dimensional floating-point vector representing the robot's joint position state. The action part is a 6-dimensional floating-point vector representing the target joint positions. The prompts in the dataset are carefully designed with three types: 75% are "default" prompts (explicitly mentioning the celebrity's name), 15% are "reference-only" prompts (only indicating the reference image), and 10% are "counterfactual" prompts (the name in the textual prompt intentionally mismatches the visible portrait), which are designed to guide models to learn to follow the reference photograph rather than the textual name. The entire dataset contains 933 unique task instruction strings. The data augmentation pipeline integrates multiple advanced computer vision models and techniques: GroundingDINO is used to detect portraits, SAM 2.1 performs mask cross-frame propagation, facial-aware corner rearrangement and Lanczos warping are combined to replace the original portrait regions with images from a dataset of 192 celebrity photos crawled from Wikimedia and Bing. Reinhard color transfer and alpha feathering blending are applied to maintain consistent scene lighting, and finally, InsightFace ArcFace is used for identity verification (cosine similarity ≥ 0.4) to ensure the quality of the augmented data.
数据集概述:so101_eval3_all
基本信息
- 数据集名称:so101_eval3_all
- 许可证:Apache-2.0
- 任务类别:机器人技术(Robotics)
- 数据规模:1M < n < 10M
- 数据格式:Parquet 文件
任务描述
机器人手臂(SO-101)需要抓取一个空的330毫升可乐罐,并将其放置在指定知名人物的印刷肖像上。每个场景中包含3张呈半圆形排列的肖像,目标人物通过参考照片和姓名提示共同指定。
数据集组成
数据规模
- 总片段数:4,195 个
- 基础遥操作片段:178 个(涉及3位标准知名人物,6种布局排列)
- 增强变体:4,017 个(每个基础片段×25种重绘变体)
覆盖范围
- 独特知名人物:192 位
- 独特任务字符串:933 个
提示类型分布
- 默认提示(75%):例如
"把可乐放在<姓名>的照片上" - 仅参考提示(15%):例如
"将罐子放在参考图像中所示的名人照片上" - 反事实提示(10%):提示中的姓名与可见肖像故意不匹配,学习逻辑应遵循参考照片而非姓名
数据结构
| 特征 | 形状 | 说明 |
|---|---|---|
observation.images.camera1 |
(480, 640, 3), H.264 30 fps | 腕部摄像头RGB图像 |
observation.images.reference |
(480, 480, 3), H.264 30 fps | 固定帧的目标名人参考照片(图像提示) |
observation.state |
(6,) float32 | 关节位置:shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper |
action |
(6,) float32 | 目标关节位置(与state相同约定) |
数据增强流程
- GroundingDINO 检测每个基础片段第0帧中的三张肖像
- SAM 2.1 在所有538帧中传播肖像掩码
- 人脸感知角点重排+Lanczos扭曲 将每个肖像区域替换为不同名人的照片(从192位名人库中采样)
- Reinhard颜色转移+羽化混合 保持与场景的光照一致性
- InsightFace ArcFace 验证替换后的人脸身份余弦相似度≥0.4
参考照片编码
每个片段的参考照片编码为恒定帧的H.264 mp4文件,使模型能够将其作为第二摄像头流与腕部摄像头图像一起处理。




