Abstract In recent years, we have witnessed the proliferation of so-called collaborative robots or cobots, that are designed to work safely along with human operators. These cobots typically use the “program from demonstration” paradigm to record and replay trajectories, rather than the traditional source-code based programming approach. While this requires less knowledge from the operator, the basic functionality of a cobot is limited to simply replay the sequence of actions as they were recorded. In this paper, we present a system that mitigates this restriction and learns to grasp an arbitrary object from visual input using demonstrated examples. While other learning-based approaches for robotic grasping require collecting a large amount of examples, either manually or automatically harvested in a real or simulated world, our approach learns to grasp from a single demonstration with the ability to improve on accuracy using additional input samples. We demonstrate grasping of various objects with the Franka Panda collaborative robot. We show that the system is able to grasp various objects from demonstration, regardless their position and rotation in less than 5 min of training time on a NVIDIA Titan X GPU, achieving over 90% average success rate.

中文翻译：

---

通过演示学习机器人掌握

摘要 近年来，我们目睹了所谓的协作机器人或协作机器人的激增，它们旨在与人类操作员一起安全地工作。这些协作机器人通常使用“演示程序”范式来记录和重放轨迹，而不是传统的基于源代码的编程方法。虽然这需要操作员的知识较少，但协作机器人的基本功能仅限于简单地重播记录的动作序列。在本文中，我们提出了一个减轻这种限制的系统，并使用演示示例学习从视觉输入中抓取任意对象。而其他基于学习的机器人抓取方法需要收集大量示例，无论是在真实世界还是模拟世界中手动或自动收集，我们的方法从单个演示中学习掌握，并能够使用额外的输入样本提高准确性。我们展示了 Franka Panda 协作机器人对各种物体的抓取。我们表明，该系统能够在 NVIDIA Titan X GPU 上不到 5 分钟的训练时间内从演示中抓取各种物体，无论它们的位置和旋转如何，平均成功率超过 90%。