SBIR Phase I: Dense, Socially-Compliant, Autonomous Delivery Robot

这个小企业创新研究第一阶段项目的更广泛的影响/商业潜力是使自主移动的机器人（AMR）能够以安全和社会兼容/可接受的方式在密集拥挤的空间中运行。一个关键的潜在成果是开发基于深度强化学习（DRL）的碰撞避免方法。 这种方法将能够处理密集的人群，并优化运行在紧凑和节能的嵌入式处理器上。这种能力将增加商业潜力，并增加对基于学习的导航方法的采用，这些方法已证明具有出色的防撞和噪音处理能力。该技术可以通过在机场，零售，医疗保健和酒店行业部署AMR来释放商业机会，这些行业的环境高度密集和动态。机场行业可能会从AMR中获得积极的影响，AMR可以在不允许使用全球定位系统（GPS）的复杂室内环境中导航，通过提供食品，饮料，这个小型企业创新研究（SBIR）第一阶段项目研究了一种混合防撞方法，使自主移动的机器人（AMR）在密集人群中安全运行，同时在稀疏场景中遵守社会规范。初步研究表明，基于深度强化学习（DRL）的方法可以用不准确、不确定的感知数据计算无碰撞机器人速度。拟议的DRL为基础的模型将被实施为一个优化的神经网络，工作在功率和成本效益的嵌入式处理器。 这项技术的关键技术障碍是：在仿真中训练的DRL模型在现实环境中可能表现不佳（称为模拟到真实的差距），由于用于在嵌入式处理器上运行的参数数量较少，与公司当前的DRL模型相比，完全训练的DRL模型可能会有一些性能下降，并且当AMR由于遮挡而导航通过密集人群时，定位模块可能计算错误的位置。第一阶段的主要目标是应对这些挑战。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值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 Phase I project is to enable autonomous mobile robots (AMRs) to operate in densely crowded spaces in a safe and socially compliant/acceptable manner. A key potential outcome is the development of a collision avoidance method based on Deep Reinforcement Learning (DRL). This method would be capable of handling dense crowds and optimized to run on compact and power-efficient embedded processors. Such abilities would increase the commercial potential and adoption of learning-based navigation methods that have demonstrated excellent collision avoidance and noise handling capabilities. The technology may unlock commercial opportunities by deploying AMRs in the airport, retail, healthcare, and hospitality industries, where the environments are highly dense and dynamic. The airport industry may derive postive impacts from AMRs that can navigate in complex, indoor environments where global positioning systems (GPS) are not allowed by providing contactless deliveries of food, beverages, and other retail products to travelers at the gate.This Small Business Innovation Research (SBIR) Phase I project investigates a hybrid collision avoidance approach enabling autonomous mobile robots (AMRs) to operate safely in dense crowds, while being socially-compliant in sparse scenarios. Preliminary research has shown that Deep Reinforcement Learning (DRL)-based approaches can compute collision-free robot velocities with inaccurate, uncertain perception data. The proposed DRL-based model will be implemented as an optimized neural network that works on power and cost-efficient embedded processors. The key technical hurdles in this technology are: the DRL model trained in simulation may not perform well in real-world environments (known as sim-to-real gap), the fully-trained DRL model may have some performance degradation compared to the company’s current DRL models due to the lower number of parameters used to run on embedded processors, and the localization modules could compute erroneous locations when the AMR is navigating through a dense crowd due to occlusions. The key objectives of Phase I are to address these challenges.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.