ASU TableTop Manipulation

培训以语言为条件的政策通常耗时且资源密集。此外，由此产生的控制器是针对他们训练过的特定机器人量身定制的，因此很难将它们转移到其他具有不同动力学为了应对这些挑战，我们提出了一种称为分层模块化的新方法高效的训练以及随后在不同类型的机器人之间传递这样的策略。该方法结合了Super-vised Attention，通过实现功能构建块的重用，弥合了模块化和端到端学习之间的差距。扩展层次结构以包括新任务，并引入了用于合成大量新颖对象的自动化管道。我们通过广泛的模拟和真实世界的机器人操作来证明这种方法的有效性实验

Language-conditioned policy training is typically time-consuming and resource-intensive. Furthermore, the resulting controllers are tailored to the specific robot they were trained on, making it difficult to transfer them to other robots with different dynamics. To address these challenges, we propose a novel approach called hierarchical modularity for efficient training and subsequent transfer of such policies across different types of robots. This method combines Supervised Attention, bridging the gap between modularity and end-to-end learning by enabling reuse of functional building blocks. We extend the hierarchy to include new tasks, and introduce an automated pipeline for synthesizing a large number of novel objects. We demonstrate the effectiveness of this approach through extensive simulation and real-world robotic manipulation experiments.

搜集汇总

背景与挑战

背景概述

该数据集旨在解决机器人学习中语言条件政策训练耗时且难以跨机器人转移的问题，提出了一种分层模块化方法，通过监督注意力实现功能构建块重用，以弥合模块化与端到端学习之间的差距。该方法支持任务扩展和自动化对象合成，并通过模拟和真实实验验证了有效性。