Collaborative Research: NeTS: Small: Reliable Task Offloading in Mobile Autonomous Systems Through Semantic MU-MIMO Control

自动驾驶汽车和无人机等移动自主系统（MASs）在城市交通、精准农业和远程监控等关键应用中发挥着关键作用。为了完成他们的任务，MASs越来越依赖于高吞吐量低延迟的计算机视觉任务流（例如，对象检测）到边缘服务器。然而，诸如阻塞、拥塞和衰落等短暂的环境因素可能会不规律地中断到边缘服务器的任务流。现有的工作分别解决了MASs任务卸载的计算和通信问题，这必然导致次优解。事实上，任务的准确性不可避免地与发送到边缘的多媒体数据的质量联系在一起，而这又取决于所采用的无线策略。然而，所使用的无线参数取决于所发送数据的质量（压缩越多，延迟越高），这最终会影响所需的任务精度。因此，为了在不影响任务准确性的情况下实现“设计弹性”应用程序，多媒体数据的语义必须从整体上和根本上与无线传输的实时优化交织在一起。该项目的核心进展是在资源有限的无线系统中设计和实验评估基于硬件的语义驱动的多媒体压缩策略和MU-MIMO传输联合优化的基本新技术。私立学校将利用该项目的支持，让少数族裔和代表性不足的学生参与研究和推广活动。作为该项目的一部分，研究生将在机器学习、嵌入式系统和无线网络的十字路口发展独特的专业知识。该项目的关键技术工作将集中在设计新颖的基于深度强化学习（DRL）的策略上，该策略将控制如何压缩采集的数据流，并通过MU-MIMO无线传输到边缘服务器。pi将利用基于分割计算的技术来避免由于压缩和MU-MIMO信道状态信息（CSI）反馈而增加的计算开销，同时保持任务精度接近原始。作为该项目的一部分，将开发基于FPGA实时处理和边缘计算的定制软件定义无线电（SDR）接口的成熟无人机原型。大规模数据收集活动将通过东北大学的64天线SDR试验台、加州大学欧文分校的无人机实验试验台和AERPAW PAWR平台进行，以：(1)收集必要的无线/多媒体数据来训练我们的算法；（ii）进行广泛的测试和性能评估。该奖项反映了美国国家科学基金会的法定使命，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。

Mobile autonomous systems (MASs) such as self-driving vehicles and drones have a pivotal role in critical applications such as urban mobility, precision agriculture and remote surveillance. To achieve their tasks, MASs increasingly rely on high-throughput low-latency streaming of computer vision tasks (e.g., object detection) to edge servers. However, ephemeral environmental factors such as blockages, congestion and fading may erratically interrupt the flow of tasks to the edge servers. Existing work has addressed computation and communication issues of task offloading by MASs separately, which necessarily leads to suboptimal solutions. Task accuracy, indeed, is inevitably tied to the quality of the multimedia data being sent to the edge, which in turns depends on the adopted wireless strategy. However, the wireless parameters being used depend on the quality of data being sent (the more compression, the higher the latency), which ultimately impacts the desired task accuracy. Thus, to achieve applications that are “resilient-by-design" without compromising task accuracy, the semantics of the multimedia data must be holistically and fundamentally intertwined with real-time optimization of wireless transmissions. The core advance of this project is the design and experimental evaluation of fundamentally novel techniques for hardware-based semantic-driven joint optimization of multimedia compression strategies and MU-MIMO transmissions in the context of resource-limited wireless systems. The PIs will leverage the support of this project to involve minority and underrepresented students in research and outreach activities. As part of the project, graduate students will develop unique expertise at the crossroads of machine learning, embedded systems and wireless networks.The key technical efforts of this project will focus on the design of novel deep reinforcement learning (DRL)-based strategies that will control how the acquired data stream is compressed and wirelessly transmitted to the edge servers through MU-MIMO. The PIs will utilize techniques based on split computing to avoid increasing computational overhead due to the compression and MU-MIMO channel state information (CSI) feedback, while keeping the task accuracy close to the original. A full-fledged drone-based prototype based on customized software-defined radio (SDR) interfaces based on FPGA real-time processing and edge computing will be developed as part of the project. Large-scale data collection campaigns will be performed with a 64-antenna SDR testbed at Northeastern, a drone experimental testbed at UC Irvine, and the AERPAW PAWR platform to (i) collect the necessary wireless/multimedia data to train our algorithms; (ii) perform extensive testing and performance evaluation.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.