danilodjor/SO101-eval2-merged-25prompts

SO101-eval2-merged是一个机器人学数据集，专门设计用于评估组合指令任务，特别是香蕉入碗任务。该数据集通过合并六个不同碗排列（gbr、rgb、bgr、brg、grb、rbg）的录制数据构建，并重新标记每个片段，使用27个提示词汇（包括原始25个提示和两个新增的Tier-1提示），以同时解决视觉颜色接地和语言覆盖问题。视觉颜色接地确保每个提到颜色的提示在所有六种碗排列中都有表示，防止模型记忆固定颜色位置；语言覆盖确保每个录制轨迹与所有解析到相同目标槽位的提示短语配对，使模型能在多种表面形式中看到相同的动作。数据集包含1518个片段，671,081帧，约2.2 GB大小，结构包括数据文件（.parquet格式）、视频文件（.mp4格式）和元数据。它旨在为模型训练和评估提供全面的组合指令数据，支持机器人视觉语言任务的研究。Single LeRobot v3.0 dataset built by merging the six per-permutation `csacan/SO101-eval2-{gbr,rgb,bgr,brg,grb,rbg}` recordings and relabelling every episode with a 27-prompt vocabulary (the 25 brg-25prompts entries plus two Tier-1 additions that fix a (colour, direction) asymmetry in the original relative family). The goal is to give Eval 2s compositional banana-in-bowl task both halves of its problem in one place: visual colour grounding — every prompt that mentions a colour is represented across all six bowl orderings, so the model cannot memorise blue = slot 0; and language coverage — every recorded trajectory is paired with every prompt phrasing that resolves to the slot it actually targets, so the model sees the same motion grounded in many different surface forms.