Brain-inspired visually guided grasping system

预计未来十年，先进机器人领域将大幅增长，主要是通过从高度可预测的工厂环境扩展到非结构化和新颖的环境。该领域的一个关键挑战是使机器人能够可靠地操纵不熟悉的物体。该项目的目标是开发一种先进的抓取系统，该系统可以感知物体的形状，并适当地定位机器人抓取器以可靠地抓取物体。这个过程对人类来说是快速而毫不费力的，但在机器人技术中却很难实现。在传统的机器人抓取中，一个人对机器人执行的精确动作进行编程。这只在装配线等场景中有用，在这些场景中，可以提前准确地知道物体的形状，以及它们在什么时候会出现在哪里。较少控制的环境要求机器人自动做出这样的判断并选择合适的抓取参数。该项目将首先采用现有的最先进的深度学习系统，该系统最近显示出自动抓取（在某些受限的情况下）的前景，在头六个月内生产出一个可工作的端到端系统。然后，该系统将被扩展到对施加的握力做出更明智的决定。在第二年，该系统的两个新版本将被开发和测试，目标是让机器人从更广泛的角度接近物体。这将允许更好的抓取质量，以及抓取更广泛的目的。最后，受深层网络和大脑之间某些相似之处的启发，我们将利用神经科学文献中的最新研究，系统地比较我们每个网络中的网络活动与猴子大脑中与抓取相关区域的活动。最后，该项目的第三年将开发一个系统的迭代，该系统将从早期网络的表现中吸取教训，并与灵长类动物的大脑进行比较。由此产生的系统将成为应用脑研究的关键产品。更广泛地说，它将为加拿大机器人行业贡献先进技术，并为HQP在这个快速增长的领域提供有价值的培训。

The field of advanced robotics is expected to grow dramatically over the next decade, primarily by expanding from highly predictable factory settings to unstructured and novel situations. A key challenge in this area is enabling robots to reliably manipulate unfamiliar objects. The project goal is to develop an advanced grasping system that can sense the shape of an object, and position a robotic gripper appropriately to reliably grasp the object. This process is rapid and effortless for humans, but it has been difficult to achieve in robotics. In conventional robotic grasping, a human programs the exact movements that the robot performs. This is only useful in scenarios such as assembly lines, where it is known in advance exactly what shapes objects have, and where they will be at what time. Less-controlled environments require robots to make such determinations and choose appropriate grasp parameters automatically. The project will begin by adapting an existing state-of-the-art deep-learning system that has shown promise recently for automatic grasping (in somewhat restricted scenarios), to produce a working end-to-end system in the first six months. The system will then be extended to make more intelligent decisions about the applied grip force. In the second year, two new variations of the system will be developed and tested, with the goal of allowing the robot to approach objects from a wider variety of angles. This will allow better-quality grasps, as well as grasping for a wider variety of purposes. Finally, inspired by certain parallels between deep networks and the brain, we will systematically compare network activity in each of our networks to activity in the grasping-related areas of the monkey brain, using recent studies in the neuroscience literature. Finally, the third year of the project will develop an iteration of the system that incorporates lessons learned from the performance of earlier networks and comparisons with the primate brain. The resulting system will become a key product of Applied Brain Research. More broadly, it will contribute advanced technology to the Canadian robotics sector, and provide valuable training for HQP in this rapidly growing area.