Multi-Robot Manipulation Planning for Forceful Manufacturing Tasks

想象一下，从五金店带着木板回家，和你的朋友一起为自己制作一张桌子。您将需要协作执行操作，如切割零件，插入钉子，钻孔和拧紧紧固件。该项目的目标是让机器人执行类似的制造任务。要做到这一点，机器人团队需要决定如何抓住工件（例如木板）以及如何移动来执行这些操作。在这个项目中，我将开发规划算法来做出这样的决定。当规划算法做出这些决策时，它们需要解决几个问题。首先，算法必须解决几何问题。假设你想拿着一块木板，让你的朋友在它的特定表面钻一个洞。你应该让你的朋友需要钻孔的表面看起来远离你（而不是让面板看起来朝向你），这样你的朋友就可以把她的身体放在你对面，舒服地钻孔。当你这样做时，你正在解决一个几何问题：你抓住工件并定位你的身体，这样你的合作者将有必要的空间来定位她自己的身体并到达工件以执行操作。协作机器人必须解决同样的问题：它们必须抓住工件并定位自己，以便它们都能到达工件并在不相互碰撞的情况下对其进行操作。随着执行操作所需的机器人数量的增加，解决几何问题变得越来越困难。其次，算法必须解决由于施加在工件上的力而引起的稳定性问题。按照上面的例子，你可能会握住木板，这样你就会把你的手掌紧紧地放在靠近钻孔点的地方，以便能够抵抗在操作过程中产生的力。首先，这需要在行动开始之前预测将要出现的力量的方向和大小。它还需要规划工件上接触点的组合和手臂的配置，这样你就能够抵抗这些力。机器人计划者也必须解决这些问题。一个特别的挑战是在解决上述几何问题的同时解决它们。最后，算法还必须解决顺序问题，因为制造一个完整的产品需要的不仅仅是一次操作。相反，机器人将需要执行多个顺序操作，例如一个接一个地钻多个孔，然后切割一块，然后插入紧固件。计划者可以天真地选择分别解决每个操作。然而，这意味着每个操作都有自己的一组抓手和几何位置，彼此独立规划。这将需要在每次操作后松开工件，并为下一次操作重新抓取和移动工件。另一方面，如果算法能够前瞻性地进行规划，例如，如果它们能够找到符合几何和稳定性约束的抓取配置，不仅适用于下一个即时操作，而且适用于随后的操作，那么最终计划的执行效率将大大提高，并且机器人可以避免冗余的解抓取/重抓取操作。然而，同时规划多个操作会使问题变得更加困难，因为在规划过程中必须更早地考虑未来操作的几何和稳定性约束。该项目的主要目标是开发一个解决所有这些限制的规划框架。提议的工作还包括建立一个多机器人系统来测试我们的算法。这个多机器人系统将从一堆制造材料开始，并执行切割、钻孔和紧固等操作来制造产品。

Imagine coming back home from a hardware store with planks of wood and working with your friend to manufacture a table for yourself. You will need to collaborate to perform operations such as cutting parts off, inserting nails, drilling holes, and screwing in fasteners. The goal of this project is to get robots to perform similar manufacturing tasks. To do this, a robot team will need to decide how to grasp the workpieces (e.g. wood planks) and how to move to perform these operations. Within this project I will develop planning algorithms that will make such decisions.When the planning algorithms make these decisions, they will need to solve several problems. First, the algorithms must solve geometric problems. Suppose you want to hold a wooden panel such that your friend will drill a hole on a particular surface of it. You would hold the panel such that the surface your friend needs to drill is looking away from you (as opposed to holding the panel such that the surface is looking towards you), so that your friend can position her body across from you and drill comfortably. As you do this you are solving a geometric problem: you grasp the workpiece and position your body such that your collaborator will have the necessary space to position her own body and reach the workpiece to perform the operation. Collaborating robots must solve the same problem: they must grasp workpieces and position themselves such that they can all reach the workpiece and perform operations on it without colliding into each other. As the number of robots required to perform an operation increases, solving the geometric problem becomes harder and harder. Second, the algorithms must solve stability problems due to forces applied on the workpiece. Keeping with the example above, you will probably hold the wooden panel such that you will rest your palms firmly against it close to the point of drilling to be able to resist the forces arising during the operation. This, first of all, requires predicting the direction and magnitude of the forces that will arise before the operation even starts. It also requires planning combinations of contact points on the workpiece and the configurations of your arms, such that you will be able to resist these forces. A robotic planner must solve these problems as well. A particular challenge is solving them simultaneously with the geometric problems mentioned above.Finally, the algorithms must also solve sequential problems because manufacturing a complete product takes much more than a single operation. Rather the robots will need to perform multiple sequential operations such as drilling multiple holes one after the other, then cutting a piece off, and then inserting fasteners. A planner can naively choose to solve each operation separately. However, this would mean that every operation has its own set of grasps and geometric position, planned independently from each other. This would require un-grasping the workpiece after every single operation, and re-grasping and moving it for the next operation. On the other hand, if the algorithms can plan with foresight, e.g. if they can find grasp configurations which obey the geometric and stability constraints not only for the next immediate operation but also for the operations following it, then the final plan would be much more efficient to execute and the robots can avoid the redundant un-grasp/re-grasp operations. Planning multiple operations simultaneously, however, makes the problem even harder because the geometric and stability constraints of future operations must be considered earlier during planning.The primary goal of this project is to develop a planning framework solving all these constraints. The proposed work also involves building a multi-robot system to test our algorithms. This multi-robot system will start with a pile of manufacturing materials and perform operations such as cutting, drilling, and fastening to build products.