PFI:AIR - TT: High-Reliability Robot Grasping for Per-Item Distribution

这个PFI：空气技术翻译项目的目的是对机器人抓取系统进行概念验证演示，错误率为10,000次中有1次或更高。在之前由美国国家科学基金会资助的工作中，研究人员开发了功能强大、成本低的机械手，当与标准工业机器人手臂和视觉系统配对时，它可以抓取足够广泛的物品，对电子商务订单履行和自动进货应用有用。然而，预期的应用需要极低的错误率。该项目将(1)开发一个系统，在现实的各种潜在客户库存上独立测试抓取系统的可靠性，以及(2)实施和评估系统在抓取过程中检测和纠正错误的新机制。到目前为止，已发表的对GRASH系统成功的评估仅限于几十到几百次试验-通常会导致1%-10%的错误率。提出的抓取误差评估方法代表了在机器人手走出实验室和工厂进入家庭、医院和车间等各种现实世界环境时验证其性能的基本能力。此外，该项目将开发独立评估抓取错误的新方法，以及提高抓取可靠性的新方法。拟议的系统将产生非常大的数据集，将在今后的工作中用于GRACH控制的数据驱动学习以及在线错误检测和纠正。此外，参与该项目的博士后将通过努力定义客户需求和开发应用驱动技术，获得创业和技术翻译经验。高可靠性的自动抓取系统承诺降低成本，提高仓储和物流业务的生产率，这是一个快速增长的经济部分。通过从自动存储和检索系统中自动选择库存，可以在较少错误来源的情况下完成订单挑选、审计和打包等任务，这在处理制药和电子产品等高价值产品时至关重要。用于评估不同对象集的抓取成功的系统和方法的开发将为在结构较少的环境中进行抓取的许多迫在眉睫的现实世界应用奠定必要的基础。

This PFI:AIR Technology Translation project is aimed at a proof-of-concept demonstration of a robotic grasping system with an error rate of 1 in 10,000 trials or better. In prior NSF-funded work, the researchers developed highly capable, low-cost robot hands which, when paired with a standard industrial robot arm and vision system, can grasp a sufficiently wide range of items to be useful for e-commerce order fulfillment and automated restocking applications. The intended applications, however, require extremely low error rates. This project will (1) develop a system to independently test the reliability of the grasping system on a realistic variety of potential customer inventory, and (2) implement and evaluate new mechanisms for the system to detect and correct errors during grasping.The proposed in-depth evaluation of the reliability of grasping systems in real-world tasks is unprecedented. To date, published assessments of grasp system success is limited to a few dozen to a few hundred trials - and typically results in error rates of 1-10%. The proposed methods for evaluation of grasp errors represents an essential capability for validating the performance of robot hands as they move out of laboratories and factories into diverse real-world settings like homes, hospitals, and shop floors. In addition, this project will develop new methods for independently assessing errors in grasping, as well as new methods for enhancing the reliability of grasping. The proposed system will produce very large data sets that will be exploited in future work for data-driven learning of grasp control and on-line error detection and correction. In addition, post docs involved in this project will receive entrepreneurship and technology translation experiences through working to define customer needs and developing application-driven technology.High-reliability automated grasping systems promise to reduce costs and enhance the productivity of the warehousing and logistics businesses that are a rapidly-growing segment of the economy. By automating the selection of inventory from automated storage and retrieval systems, tasks such as order picking, auditing and packing can be accomplished with fewer sources of error, which is crucial in handling high-value products such as pharmaceuticals and electronics. The development of systems and methods for assessing grasping success for diverse object sets will lay the essential groundwork for many imminent real-world applications of grasping in less structured environments.