RI: Small: A Novel Framework for Informed Manipulation Planning

人类使用多种类型的操作来完成日常任务，并轻松地排序和执行相关操作。机器人被限制在非常简单的任务上，这些任务往往被操作这些机器人的人煞费苦心地分解到令人痛苦的细节水平。今天，比以往任何时候都更迫切地需要开发能够像人类一样无缝地对操作进行推理的机器人。操作规划查找将机器人及其交互对象从开始状态移动到目标状态的一系列动作。在这样做的过程中，形成并打破了联系，施加了限制，并发生了一些高级推理：抓取、放置、重新抓取和重新排列物体的情况经常发生。将操作规划与经典的人工智能规划和正式方法相结合，将产生能够打扫医院房间、在ICU执行任务或帮助康复患者的机器人所期望的那种行为。该项目倡导一种新颖的、统一的操作规划框架。它采用以约束为中心的观点，特别是使用定义低维子空间或模式的流形约束，机器人必须在这些子空间或模式之间转换。转换的定义也是以约束为中心的，由定义模式的约束的组合定义，并且仅在用于考虑模式的统一方法中是可能的。这项工作依赖于微分几何的构造和多分辨率搜索方案的使用。首先，该项目将推广基于采样的机器人运动规划算法，将规划策略与操作中经常发现的约束满足程度解耦。其次，在搜索过程中，将定义和利用模式之间转换图的概念，以自动发现模式之间可行的转换并生成操作计划。这项工作将从一种特定但一般的约束类型开始，即流形约束，然后扩展到其他类型。建议的研究将调查建议的框架提供的最优化和健壮性保证，并将确定在操作规划中使用约束作为统一结构的限制。这项工作将得到模拟和现实场景中广泛实验的支持，这些场景涉及几个对象、橱柜、货架、一个移动机械手和两个固定的最先进的机械手。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Humans use many types of manipulation to accomplish daily tasks and easily sequence and execute pertinent actions. Robots are confined to very simple tasks that are often painstakingly broken down to an excruciating level of detail by humans who operate those robots. Today, more than ever, there is an urgent need to develop robots that reason about manipulation as seamlessly as humans do. Manipulation planning finds a sequence of actions that move the robot and the objects it interacts with from a start state to a goal state. In doing so, contacts are formed and broken, limits are imposed, and some high-level reasoning occurs:. Grasping, placing, regrasping, and rearranging objects often take place. The combination of manipulation planning with classical AI planning and formal methods will lead to the kind of behavior that is expected from a robot capable of cleaning a hospital room, performing tasks in an ICU, or assisting a recovering patient.This project advocates a novel, unifying framework for manipulation planning. It adopts a constraint-centric view, in particular using manifold constraints which define lower-dimensional subspaces, or modes, among which the robot must transition. The definition of transitions is also constraint-centric, defined by the combination of the constraints which define modes, and is only possible within the unified approach used to consider modes. The work depends on constructs from differential geometry and the use of multi-resolution search schemes. First, the project will generalize sampling-based algorithms for robot motion planning by decoupling the planning strategy from the satisfaction of constraints often found in manipulation. Second, the concept of a transition graph between modes will be defined and exploited during search to automatically find viable transitions between modes and produce manipulation plans. The work will start with a specific but general type of constraint, manifold constraints, and later expand to other types. The proposed research will investigate optimality and robustness guarantees provided by the proposed framework and will identify the limits of using constraints as a unifying construct in manipulation planning. The work will be supported by extensive experimentation in both simulation and realistic scenarios that involve several objects, cabinets, shelves, a mobile manipulator, and two stationary state-of-the-art manipulators.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.