SBIR Phase I: Advanced Artificial Intelligence for Robotic E-Commerce Pick-and-Pack Automation

小型企业创新研究(SBIR)第一阶段项目的更广泛影响/商业潜力是促进可靠、灵活和可扩展的物流网络的发展，以分发基本物品和用品。最近的预测表明，到2022年，美国电子商务的总销售额将超过9000亿美元。整个美国电子商务中的件处理自动化市场估计为96亿美元。将商品从生产商运往消费者的过程涉及许多接触点，运营商挑选和包装单个商品。这些过程目前都是手工进行的，重复性很高，导致了很高的伤害率。这些过程中的错误对电子商务提供商来说代价高昂，并可能导致关键供应损失或严重延误。然而，由于仓库流程的显著多样性，自动化拣选和包装一直是具有挑战性的，需要新的人工智能(AI)机器人控制系统来操作仓库中的大量独特物品。用于机器人部署的人工智能操作系统的创新可以积极影响国家供应链的方方面面，并确保在美国境内快速而强大的必需品和消费品分销网络。这个小企业创新研究(SBIR)第一阶段项目推动了机器人挑选从模拟到现实的迁移学习的转换。通过生成数百万个模拟的机器人抓取和传感器读数，深度神经网络可以被训练成可靠地为特定应用挑选和放置各种对象。该项目将开发和评估机器人拾取的新算法，以开发灵活的机器人控制软件，用于跨越各种物理实例的材料处理。研究目标是减少抓取策略的计算时间，同时规划跨多个工具的抓取，并将抓取策略与电子商务配送中心遇到的订单处理过程相结合。研究目标将在一组测试对象上通过标准化的机器人挑选系统进行系统测试，以评估性能。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to advance the development of a reliable, flexible, and scalable logistics network for distributing essential items and supplies. Recent projections suggest that by 2022, total US e-commerce sales will exceed $900 B. The market for piece-handling automation across US e-commerce is estimated at $9.6B.The process for getting items from producer to consumer involves many touchpoints where operators pick and pack individual items. These processes are currently manual and highly repetitive, incurring a high rate of injuries. Errors in these processes are costly to e-commerce providers and may result in critical supplies getting lost or significantly delayed. However, automating pick-and-pack has been challenging due to significant diversity in warehouse processes and requires new Artificial Intelligence (AI) robotic control systems that can manipulate a large number of unique items in warehouses. Innovations in AI operating systems for robotic deployments can positively impact all aspects of the national supply chain and ensure a rapid and robust distribution network of essential items and consumer goods within the United States.This Small Business Innovation Research (SBIR) Phase I project advance the translation of simulation-to-reality transfer learning for robotic picking. By generating millions of simulated robotic grasps and sensor readings, deep neural networks can be trained to reliably pick and place a wide variety of objects for a particular application. This project will develop and evaluate new algorithms for robotic piece picking to develop flexible robotic control software for material handling across a variety of physical instantiations. The research objectives are to decrease computation time for grasping policies, plan grasps across multiple tools simultaneously, and integrate grasp policies with order handling processes encountered in e-commerce distribution centers. The research objectives will be systematically tested on a standardized robotic picking system on a set of test objects to evaluate performance.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.