BUDS

我们通过技能发现来解决现实世界中的长视野机器人操作任务。我们提出了一种自下而上的方法，从未分段的演示中学习可重用技能库，并使用这些技能来综合长期的机器人行为。我们的方法首先通过凝聚聚类从每个演示构建分层任务结构。从多任务演示的任务结构中，我们根据重复模式识别技能，并通过分层模仿学习训练目标条件的感觉运动策略。最后，我们训练一个元控制器来组合这些技能来解决长视野操作任务。整个模型可以在 30 分钟内收集的一小组人类演示上进行训练，无需进一步注释，从而可以修改为现实世界的部署。我们在模拟环境和真实机器人上系统地评估了我们的方法。我们的方法在多阶段操作任务中表现出优于最先进的模仿学习方法的性能。此外，与从单个任务中发现的技能相比，从多任务演示中发现的技能将平均任务成功率提高了 8%。

We address real-world long-horizon robotic manipulation tasks via skill discovery. We propose a bottom-up approach to learn a reusable skill library from unsegmented demonstrations, and use these skills to synthesize long-term robotic behaviors. Our method first constructs a hierarchical task structure from each demonstration via agglomerative clustering. From the task structures of multi-task demonstrations, we identify skills based on recurring patterns, and train goal-conditioned sensorimotor policies via hierarchical imitation learning. Finally, we train a meta-controller to compose these skills to solve long-horizon manipulation tasks. The entire model can be trained on a small set of human demonstrations collected within 30 minutes, without further annotations, enabling modification for real-world deployment. We systematically evaluate our method in both simulated environments and on real robots. Our method outperforms state-of-the-art imitation learning methods in multi-stage manipulation tasks. Furthermore, compared with skills discovered from single tasks, skills derived from multi-task demonstrations improve the average task success rate by 8%.

搜集汇总

背景与挑战

背景概述

BUDS是一个用于机器人学习的数据集，提出了一种自下而上的技能发现方法，从多任务演示中学习技能库，以提升长视野操作任务的性能。该方法在模拟和真实机器人上进行了评估，无需额外注释，基于少量人类演示即可训练。