CAREER: Facilitating Autonomy of Robots Through Learning-Based Control

无人机技术在过去几十年中取得了重大进展。然而，将许多不同制造商的异构无人机大规模应用于现实世界仍然是非常具有挑战性的。一个主要原因是，每当一个新的无人机是建立，无人机的规划和控制算法通常必须非常仔细地设计和无人机采取的行动通常必须通过相当大的调整努力费力编程。为了消除（如果不是减少）这些限制，这个教师早期职业发展（Career）项目建立了一个新的基于学习的框架，使无人机具有“从其他无人机的经验中学习”的新能力，尽管它们的动力和平台不同。这种无人机规划和控制的设计方法将大大减少无人机的设计，测试，评估和认证，独特而有效地为其操作环境中的应用定制。综合研究和教育活动将为纽约西部地区的学生提供无人机技术的实践经验和实习机会，以便更好地为美国无人机系统行业的未来劳动力做好准备。该项目将建立一种新的基于学习的前馈控制框架，并为无人机提供学习三种特定技能的新能力，即(1)如何生成动态可行的轨迹，(2)如何感知和补偿外部干扰，以及(3)如何从他人的学习经验中学习，称为“动态学习”。这三项技能对于无人机执行复杂任务至关重要，也是理解一个机器人如何有效地从其他具有不同动态的机器人收集的经验中学习的基础。该方法的关键是通过添加前馈学习信号来自动调整基线规划器和控制器的原始输出的架构，以提高无人机的飞行性能。这种学习框架既不是为了完全取代现有的规划和控制方法，也不是为了争夺最高的优化性能，而是为了提供一种优雅的学习机制，这种机制具有高度的适应性和合理的效率，涉及最小的硬件修改和软件重构。该项目由跨部门机器人基础研究项目支持，由工程（ENG）和计算机与信息科学与工程（CISE）联合管理和资助。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。

Drone techniques have achieved significant progress in the past decades. However, it is still very challenging to massively bring heterogeneous drones by many different manufacturers to real-world applications. One main reason is that, whenever a new drone is built, the planning and control algorithms for the drone usually have to be designed very carefully and the actions for the drone to take usually have to be laboriously programmed with considerable tuning effort. To remove, if not lessen, such limitations, this Faculty Early Career Development (CAREER) project establishes a novel learning-based framework that equips drones with new capabilities of "learning from the experience" of other drones despite their different dynamics and platforms. This approach to design of planning and control of drones will significantly reduce the design, test, evaluation and certification of drones, uniquely and efficiently customized for applications in their operating environment. The integrated research-and-education activities will provide students in the Western New York area with hands-on experience and internship opportunities on drone techniques, toward better preparing the future workforce for the unmanned aerial system industry in the United States.This project will establish a novel learning-based feedforward control framework and equip drones with new capabilities for learning three particular skills, i.e., (1) how to generate a dynamically feasible trajectory, (2) how to sense and compensate external disturbances, and (3) how to learn from others' learned experience, called "dynamic learning." These three skills are crucial for drones to perform complex tasks, and the foundation for understanding of how one robot could efficiently learn from the experiences gathered by other robots with different dynamics. Key to this approach is an architecture that automatically adjusts the original outputs of the baseline planners and controllers by adding feedforward learning signals to improve drone's flight performance. This learning framework is neither to completely replace the existing planning and control methods nor to compete for the highest optimized performance possible but rather to provide an elegant learning mechanism that is highly adaptable and reasonably efficient involving minimal hardware modification and software reconfiguration for commodity drones.This project is supported by the cross-directorate Foundational Research in Robotics program, jointly managed and funded by the Directorates for Engineering (ENG) and Computer and Information Science and Engineering (CISE).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.