BC-Z

我们研究使基于视觉的机器人操纵系统能够泛化到新任务的问题，这是机器人学习中长期存在的挑战。我们从模仿学习的角度来应对这一挑战，旨在研究如何扩展和扩大所收集的数据来促进这种泛化。为此，我们开发了一种交互式且灵活的模仿学习系统，该系统可以从演示和干预中学习，并且可以以传达任务的不同形式的信息为条件，包括预先训练的自然语言嵌入或人类执行任务的视频。

We study the problem of enabling vision-based robotic manipulation systems to generalize to novel tasks, a long-standing challenge in robot learning. We address this challenge from the perspective of imitation learning, aiming to investigate how to scale up and expand collected data to facilitate such generalization. To this end, we develop an interactive and flexible imitation learning system that can learn from demonstrations and interventions, and can be conditioned on diverse forms of task-conveying information, including pre-trained natural language embeddings or videos of humans performing tasks.

搜集汇总

背景与挑战

背景概述

BC-Z是一个面向机器人学习的数据集，专注于通过模仿学习提升机器人操纵系统的任务泛化能力，支持多种信息条件输入，数据集大小为467.6GB。