NRI: Collaborative Research: Towards Robots with Human Dexterity

尽管“硬件”和“湿软件”要慢得多，但人类的灵巧程度远远超过了现代机器人。这个项目研究的任务显然很简单——管理开门所需的手部运动的运动学约束；处理一杯咖啡中液体晃动的动态复杂性——这对机器人来说是一个巨大的挑战，但人类却能轻松完成。关键思想是，人类通过巧妙地组合不需要持续干预的原始动态动作，与这些物体进行熟练的物理交互。一种新的理论描述了这种方法的有效性，并通过对人类受试者的实验进行了测试。这一理论被用于将类似的技能转移到机器人身上，尽管机器人的硬件显然不同。如果成功，这些机器人将更有能力，更容易理解，更能与人类合作。实验的核心挑战是确定人类的基本策略。出色的物理交互任务能力。三个假设反映了当代运动神经科学的主要主题进行了测试：人类1)开发物体动力学模型，足以预先计算和执行所需的手部动作（类似于现代机器人编程）；2)选择力量和动作，使肌肉消耗最小化（类似于优化效率）；或者3)利用动态原语健壮地实现令人满意的（足够好的）性能。理论上的挑战是将大脑（或计算机）的信息处理与物理对象的“能量处理”及其相互作用结合起来，形成一个连贯的解释。经典等效电路理论被重新定义为新经典等效网络理论，将动态运动原语与机械阻抗（交互动力学）相结合。机械阻抗具有一个关键的特性，即组合性，它克服了维度的诅咒。

Despite vastly slower "hardware" and "wetware," human dexterity vastly out-performs modern robots. This project studies apparently-simple tasks - managing the kinematic constraint on hand motion required to open a door; and dealing with the dynamic complexity of liquid sloshing in a cup of coffee - that profoundly challenge robots but humans perform with ease. The key idea is that humans manage skillful physical interaction with these objects by exploiting clever combinations of primitive dynamic actions that do not require continuous intervention. A novel theory to describe the effectiveness of this approach is developed and tested by experiments with human subjects. The theory is applied to transfer comparable skill to robots, despite manifestly different hardware. If successful, these robots will be more capable, more comprehensible, and more collaborative partners with humans.The central experimental challenge is to determine the essential strategy underlying humans? remarkable competence in physical interaction tasks. Three hypotheses reflecting major themes in contemporary motor neuroscience are tested: Humans 1) develop models of object dynamics sufficient to pre-compute and execute required hand motions (similar to modern robot programming); 2) choose forces and motions to minimize muscular effort (similar to optimizing efficiency); or 3) exploit dynamic primitives to robustly achieve satisficing (good-enough) performance. The theoretical challenge is to formulate a coherent account combining the information-processing of brains (or computers) with the "energy-processing" of physical objects and their interactions. Classical equivalent circuit theory is re-purposed to define a neo-classical equivalent network theory, combining dynamic motion primitives with mechanical impedances (interactive dynamics). Mechanical impedances enjoy a key property, compositionality, that overcomes the curse of dimensionality.