NRI: Robust Stochastic Control for Agile Aerial Manipulation

一种新型的飞行机器人开始不仅导航和观察它们的周围环境，而且还可以在人类难以到达的地方接触和操纵物体。这样的系统将帮助人们在不安全或偏远的地点进行操作，并将自动化人工劳动密集型任务，如包裹递送、农业检查和基础设施修复。目前的空中机械手原型缺乏控制保真度，无法确保此类操作所期望的可靠性和效率。为了克服这些限制，拟议的项目开发了利用飞行器能力的新型控制技术。如果成功，这项研究项目将使在极端环境中灵活而安全的空中操纵成为可能，而目前使用标准方法是不可能或不可行的。这项研究的目标是为机器人实现具有内置鲁棒性的规划和控制方法，这些机器人可以在自主和人工辅助模式下与环境交互和操纵环境。这是通过将感知-控制耦合问题作为统计学习问题来实现的，并自适应地计算决策策略以优化未来的性能并将违反安全的概率降至最低。该方法的核心是一种可证明稳定的自适应控制方法，该方法配备了关于最大期望成本和碰撞概率的概率鲁棒性保证。这些界限对应于通过贝叶斯可能近乎正确的分析得出的测量集中不等。两个实验平台为以下方面提供了概念验证：1)用于重复性任务(如负载运送、作物采样和远程清洁)的自主“空气抓取器”；2)协作机器人“空中之手”，直接协助人类操作员进入危险或难以进入的地方，例如在救援或安全敏感任务期间进行极端环境下的检查和维修。实施的技术是普遍适用的，并将作为开放源码的ROS兼容软件发布。

A new class of flying robots are beginning to, not only navigate and observe, their surroundings, but also reach and manipulate objects in places that are difficult for humans to go. Such systems will assist people through manipulation in unsafe or remote locations, and will automate manual labor-intensive tasks such as package delivery, agricultural inspection, and infrastructure repair. Current aerial manipulator prototypes lack the control fidelity to ensure reliability and efficiency that is expected from such operations. To overcome these limitations, the proposed project develops novel control techniques that exploit the capabilities of the aerial vehicle. If successful, this research project will enable agile and safe aerial manipulation in extreme environments that is presently impossible or infeasible using standard methods. The goal of this research is the realization of planning and control methods with built-in robustness for robots that can interact with and manipulate the environment in autonomous and human-assisted modes. This is accomplished by posing the coupled perception-control problem as a statistical learning problem and adaptively computing decision policies to optimize future performance and minimize probability of safety violation. At the core of the approach lies a provably-stable adaptive control methodology equipped with probabilistic robustness guarantees in terms of maximum expected cost and probability of collision. These bounds correspond to concentration-of-measure inequalities derived through Bayesian probably-approximately-correct analysis. Two experimental platforms provide proof-of-concept for: 1) an autonomous "Air-gripper" for repetitive tasks such as load delivery, crop sampling, and remote cleaning; 2) co-robotic "hands in the sky" in direct assistance to a human operator enabling access to dangerous or difficult-to-access places, e.g. for inspection and repair in extreme environments, during rescue or security-sensitive missions. The implemented techniques are generally applicable and will be released as open-source ROS-compatible software.