Evaluation of: image feature quantization effects on object grasp pose estimation in a tabletop manipulation environment; of high-level planner accuracy on a prepared set of simple command permutations

A small data set that was used to test the effects of quantizing pixel-wise embedding vectors in an open-set image segmentation and spatial mapping system. Contains:rgb images of the scenesdepth images of the scenesground truth reference positions and grasp poses generated semi-automatically by human labelersinferred object grasp poses using two different mapping systems, a 2.5d approach processing the RGB-D images from a Zivid camera directly, and a 3d approach integrating multiple RGB-D images from a realsense camera into an octree occupancy map data structureAddition: a supplementary data set including the outputs from evaluating an LLM-based high-level planner on a set of natural language queries has been appended.<br>

本小型数据集用于测试逐像素嵌入向量量化（pixel-wise embedding vector quantization）在开放集图像分割与空间映射系统中的效果。数据集包含场景的RGB图像、场景的深度图像、由人工标注员半自动生成的真值参考位置与抓取位姿，以及通过两种不同映射系统推导得到的物体抓取位姿：其一为直接处理Zivid相机（Zivid camera）采集的RGB-D图像（RGB-D image）的2.5D方法，其二为将RealSense相机（RealSense camera）采集的多帧RGB-D图像整合至八叉树占用地图数据结构（octree occupancy map data structure）的3D方法。
补充说明：本次追加了一套辅助数据集，其中包含基于大语言模型（LLM/Large Language Model）的高层规划器针对一系列自然语言查询的评估输出结果。