RLBench Demonstration Dataset

该数据集包含了来自RLBench框架的多种机器人操作任务的数据，每个任务都有50个示范，包括捡起杯子、按下按钮、将垃圾放入垃圾桶、将手机放在底座上以及揭开锅盖等任务。该数据集利用RLBench内置的数据生成器生成，并包含了提取的航点，用于训练模型预测下一个关键点状态。该数据集大约涵盖了7.5千个问答对，涉及三个问答任务，旨在为机器人操作提供具身推理支持。

This dataset contains data from various robotic manipulation tasks within the RLBench framework. Each task has 50 demonstrations, including picking up a cup, pressing a button, placing trash into a trash bin, putting a phone on its base, and lifting a pot lid. Generated using RLBench's built-in data generator, this dataset also includes extracted waypoints for training models to predict next keyframe states. It encompasses approximately 7,500 question-answer pairs across three question-answering tasks, with the goal of providing embodied reasoning support for robotic manipulation.

搜集汇总

背景与挑战

背景概述

RLBench Demonstration Dataset是一个用于支持基于视觉的机器人操作学习的数据集，特别关注Q-attention和coarse-to-fine Q-attention技术。数据集需要YARR框架和RLBench 1.1.0的支持，并提供了详细的安装和实验运行指南。