ETH Agent Affordances

对于移动机器人来说，与铰接物体的交互是一项具有挑战性但重要的任务。为了应对这一挑战，我们提出了一种新颖的闭环控制管道，它将来自提供率估计的操作先验与基于采样的全身控制相结合。我们引入了代理感知提示的概念，该概念充分反映了代理的能力和体现，并且我们表明它们的性能优于仅以末端执行器几何形状为条件的最先进的对应物。此外，发现闭环提供推理允许代理将任务划分为多个非连续运动，并从故障和意外状态中恢复。最后，管道能够在现实世界中以高成功率（打开：71%，关闭：72%）执行长距离移动操作任务，即打开和关闭烤箱。

For mobile robots, interacting with articulated objects is a challenging yet critical task. To address this challenge, we propose a novel closed-loop control pipeline that combines manipulation priors that provide rate estimates with sampling-based whole-body control. We introduce the concept of agent-aware prompts, which fully reflect the agent’s capabilities and embodiment, and demonstrate that these prompts outperform state-of-the-art counterparts conditioned solely on end-effector geometry. Additionally, we find that closed-loop affordance reasoning enables the agent to decompose tasks into multiple discontinuous motions and recover from failures and unexpected states. Finally, the pipeline can execute long-distance mobile manipulation tasks, specifically opening and closing ovens, in the real world with high success rates (opening: 71%, closing: 72%).

搜集汇总

背景与挑战

背景概述

该数据集专注于移动机器人与铰接物体的交互，提出了一种结合操作先验和全身控制的闭环控制管道，并引入代理感知提示以提升性能。该方法在现实世界任务中实现了高成功率，例如开关烤箱。