Topology-based Motion Synthesis

类人机器人研究领域的主要驱动力之一是希望实现机器人与环境或人之间的密切接触的运动，例如在抬着受伤的人时，处理背包或衣服的背带等柔性物体。目前，由于开放式环境中潜在的计算复杂性，这些应用似乎超出了现有运动合成技术的能力。传统的运动合成方法存在两大瓶颈。首先，在机械臂节段和物体之间存在大量紧密接触的情况下，进行碰撞检测和避障需要大量的计算。其次，随着环境的变化，任何特定的计算解都很容易变得无效。例如，如果机器人正在处理诸如背包之类的物体，即使这种柔性物体的微小变形和物体尺寸的微小变化(例如，在空袋和填充袋之间)也可能需要以当前解决问题的方式完全重新规划。类似的问题也出现在计算机动画领域，其中需要实时控制角色-摆脱预先编程的运动的静态序列。尽管这个世界看起来更受控制，因为它是由动画设计师创建的，但实际上人们强烈希望创建游戏和模拟系统，在这些系统中，用户可以不断地与世界互动，并期望动画系统做出相应的反应。这就需要在运动合成技术上取得如上所述的同样的进步。根本问题在于对世界和机器人状态的表示。通常，运动是在以广义坐标的级别表示的完整配置或状态空间中合成的，该广义坐标列举了所有关节角度及其相对于某个世界参考系的3D位置/方向。这意味着需要在非常大的搜索空间中进行大量的碰撞检查计算和随机探索。此外，很难在这种描述级别上编码更高级别的语义规范，因为广义坐标的各个值不能告诉我们任何事情，除非进行进一步的计算以确保满足相关约束。当在大型数据库中搜索运动时，这尤其不方便。本研究的重点是通过开发利用这些问题中潜在的拓扑结构的方法来缓解这些问题，例如在姿势空间中。这允许我们定义一个新的搜索空间，其中的坐标基于拓扑关系，例如链接段之间的关系。我们用“拓扑坐标”来指代这个空间。在前期工作中，我们已经展示了这种观点对于与紧密接触的角色进行有效的运动合成的实用性。我们还证明了这种方法在对语义相似的动作进行分类时更有效。在这个项目中，我们将开发一种更通用的此类技术框架，适用于由自主仿人机器人和虚拟动画角色执行的大类任务。此外，我们将通过本田欧洲研究院和日本Namco Bandai的合作伙伴，在与行业相关的平台上实施我们的技术。

One of the major drivers of research in the area of humanoid robotics is the desire to achieve motions involving close contact between robots and the environment or people, such as while carrying an injured person, handling flexible objects such as the straps of a knapsack or clothes. Currently, these applications seem beyond the ability of existing motion synthesis techniques due to the underlying computational complexity in an open-ended environment. Traditional methods for motion synthesis suffer from two major bottlenecks. Firstly, a significant amount of computation is required for collision detection and obstacle avoidance in the presence of numerous close contacts between manipulator segments and objects. Secondly, any particular computed solution can easily become invalid as the environment changes. For instance, if the robot were handling an object such as a knapsack, even small deformations of this flexible object and minor changes in object dimensions (e.g., between an empty bag and a stuffed bag) might require complete re-planning in the current way of solving the problem. Similar issues arise in the area of computer animation, where there is a need for real-time control of characters - moving away from static sequences of pre-programmed motion. Although it may seem that this world is much more contained, as it is created by an animation designer, there is in fact a strong desire to create games and simulation systems where the users get to interact with the world continually and expect the animation system to react accordingly. This calls for the same sort of advances in motion synthesis techniques as outlined above.The fundamental problem lies in the representation of the state of the world and the robot. Typically, motion is synthesizes in a complete configuration or state space represented at the level of generalized coordinates enumerating all joint angles and their 3D location/orientation with respect to some world reference frame. This implies the need for large amounts of collision checking calculations and randomized exploration in a very large search space. Moreover, it is very hard to encode higher level, semantic, specifications at this level of description as the individual values of the generalized coordinates do not tell us anything unless further calculations are carried out to ensure satisfaction of relevant constraints. This is particularly inconvenient when searching for a motion in a large database. The focus of this research is to alleviate these problems by developing methods that exploit the underlying topological structure in these problems, e.g., in the space of postures. This allows us to define a new search space where the coordinates are based on topological relationships, such as between link segments. We refer to this space in terms of 'topology coordinates'. In preliminary work, we have shown the utility of this viewpoint for efficient motion synthesis with characters that are in close contacts. We have also demonstrated that this approach is more efficient for categorizing semantically similar motions. In this project, we will develop a more general framework of such techniques that will be applicable to a large class of tasks carried out by autonomous humanoid robots and virtual animated characters. Moreover, we will implement our techniques on industrially relevant platforms, through our collaborators at Honda Research Institute Europe GmbH and Namco Bandai, Japan.