RTML: Large: Co-design of Hardware and Algorithms for Energy-efficient Robot Learning

从昆虫大小的传单到手掌大小的卫星，微型低能耗自主机器人车辆在各种行业中具有巨大影响的潜力，包括消费电子，高带宽通信，搜索和救援行动以及太空探索，仅举几例。支持这些应用程序的下一代低能耗计算硬件必须具有适应性，即，在飞行中识别新环境，实时学习其特征，并调整其计算策略以最大限度地减少计算任务所需的能耗。该项目将帮助车辆提高感知和决策算法的准确性，只需通过实验获取不同的环境视角，并利用其运动知识到地面，并通过机器学习提高其观察能力。该项目还寻求在麻省理工学院开发新的研究生和本科课程，它将为高中生提供服务，让女性和代表性不足的群体参与进来，从而帮助培养未来的美国劳动力。该项目将开发实时机器人学习算法和硬件，重点关注三个核心领域。首先，该项目将开发实时连续机器人学习系统，通过在新环境中的快速学习来提高机器人感知和决策的性能。其次，该项目将开发实时主动机器人学习系统，以有效地在提高感知和决策算法的准确性与专注于完成手头任务之间取得平衡。第三，该项目将开发用于能量可扩展感知和决策的实时适应性机器人学习系统，该设计允许有效的准确性和能量权衡。该项目将通过启用新的低能耗机器人系统，帮助开发用于实时机器人学习的新硬件和算法。该项目还将与DARPA的一个协同项目合作进行相关硬件开发。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。

Miniature low-energy autonomous robotic vehicles, ranging from insect-size flyers to palm- size satellites, hold the potential for tremendous impact in a diverse set of industries, including consumer electronics, high-bandwidth communications, search and rescue operations, and space exploration, just to name a few. Next-generation low-energy computing hardware that will enable these applications must be adaptable, i.e., recognizing new environments on the fly, learning their characteristic features in real time, and adapting its computing strategy to minimize the energy consumption required for computing task. This project will help realize vehicles that are able to improve the accuracy of their perception and decision making algorithms, simply by experimenting with obtaining a diverse set of viewpoints of the environment and utilizing the knowledge of its motion to ground and improve its observation via machine learning. The project also seeks to develop new graduate and undergraduate courses at MIT, it will enable outreach for high school students, involve women and underrepresented groups, thus helping train the future US workforce. This project will develop real-time robot learning algorithms and hardware focusing on three core areas. Firstly, the project will develop real-time continuous robot learning systems that improve performance of robot perception and decision making by rapid learning in new environments. Secondly, the project will develop real-time active robot learning systems to efficiently decide the balance between improving accuracy of perception and decision making algorithms and focusing on accomplishing the task at hand. Thirdly, the project will develop real-time adaptable robot learning systems for energy scalable perception and decision making, where the design allows for efficient accuracy-energy tradeoffs. The project will help develop new hardware and algorithms for real-time robot learning, by enabling new low-energy robotic systems. The project will also collaborate with a synergistic DARPA program for related hardware development.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.