Learning and Adaptation for Long-Term Autonomous Robotics Applications

长期自主机器人应用的学习和适应

• 批准号: RGPIN-2014-04634
• 负责人: Schoellig, Angela
• 金额: $ 2.26万
• 依托单位: University of Toronto
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2019
• 资助国家: 加拿大
• 起止时间: 2019-01-01 至 2020-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=683944
• 关键词: Learning; Adaptation; Long; Term; Autonomous

The majority of today's robots are unable to reliably operate in unknown, changing and generally uncontrolled environments. Their systems rely on knowing in advance the specifics of every possible situation they might encounter. They are not able to adapt to new situations. This is one of the main reasons why, to date, autonomous (mobile) robots have been primarily deployed in highly controlled environments (e.g. manufacturing plants or warehouses). To overcome this major limitation, the proposed research program will develop novel robotics algorithms and systems that extend robot operations to uncontrolled environments by enabling robots to learn and adapt. Such capabilities will lead to a major paradigm shift in robotics towards the long-term deployment of autonomous robots in uncontrolled real-world application scenarios such as inspection, remote sensing, resources monitoring, and space exploration.**Assuming that useful a priori information about a robotic system and its environments is available, we focus on hybrid model-and-data-driven approaches, which aim to combine the best of two worlds: a priori model information assures safe operation initially (e.g. based on a robust model-based controller design), while based on the data collected during operation the model can be refined and the robot's performance can be gradually improved. **This approach blends the worlds of machine learning and data mining with classical control and estimation theory, and has proven to be successful in isolated scenarios. Recent work by the applicant has shown that difficult-to-model aerodynamic effects can be learned effectively and efficiently resulting in precise high-speed flight maneuvers. She has also shown that a ground robot can learn to compensate for unknown rough terrain. These results show that hybrid model-and-data-driven control is a promising new direction for robotics that has the potential to help bridge the gap from simple lab experiments to serious long-term service applications. **We seek to enable reliable long-term single- and multi-robot applications in uncontrolled environments. To do so, we focus on two long-term objectives: **(I) To develop learning algorithms and software that enable mobile aerial and ground robots to operate over the long term in challenging, uncontrolled real-world environments; and*(II) To enable technology transfer of these techniques to real-world applications by focusing on the critical technology gaps and rigorous field testing needed to make this leap.**In order to address these objectives, it will be necessary to carry out both extensive theoretical analyses and a significant experiment campaign using ground and aerial robots. The two objectives above can be further divided into several shorter-term goals to be addressed in the next five years: ** designing task-specific learning strategies for operations in uncontrolled environments;* devising methods that enable task-independent learning;* developing collaborative learning schemes for heterogeneous and homogeneous teams of robots; and* demonstrating successful learning on different robot platforms over extended periods of time.**The applicant will collaborate with Canadian corporations and organizations such as MDA, CSA, Clearpath and PrecisionHawk. These collaborations facilitate extensive testing of the algorithms on commercial aerial and ground robot platforms and provide the applicant with feedback from end users. Ultimately, based on the applicant's findings there will be tremendous opportunities for Canadian industries to develop personal, enterprise and space robotic solutions, and diversify their current portfolio into multi-billion application areas such as remote sensing, natural resources monitoring, and surveillance.

今天的大多数机器人都无法在未知的、不断变化的和通常不受控制的环境中可靠地运行。他们的系统依赖于提前了解他们可能遇到的每一种情况的细节。他们不能适应新的环境。这是迄今为止，自主（移动）机器人主要部署在高度控制的环境（例如制造工厂或仓库）的主要原因之一。为了克服这一主要限制，拟议的研究计划将开发新的机器人算法和系统，通过使机器人能够学习和适应，将机器人操作扩展到不受控制的环境。这种能力将导致机器人技术的重大范式转变，使自主机器人在不受控制的现实应用场景中长期部署，如检查、遥感、资源监测和空间探索。**假设关于机器人系统及其环境的有用先验信息是可用的，我们专注于混合模型和数据驱动的方法，其目的是结合两个世界的优点：先验模型信息最初确保安全运行（例如基于鲁棒模型的控制器设计），而基于运行过程中收集的数据可以改进模型，机器人的性能可以逐步提高。**这种方法将机器学习和数据挖掘与经典控制和估计理论相结合，并已被证明在孤立的场景中是成功的。申请人最近的工作表明，难以建模的空气动力学效应可以有效和高效地学习，从而实现精确的高速飞行机动。她还展示了地面机器人可以学会对未知的崎岖地形进行补偿。这些结果表明，混合模型和数据驱动控制是机器人技术的一个有前途的新方向，它有可能帮助弥合从简单的实验室实验到严肃的长期服务应用的差距。**我们寻求在不受控制的环境中实现可靠的长期单机器人和多机器人应用。为此，我们专注于两个长期目标：**(I)开发学习算法和软件，使移动空中和地面机器人能够在具有挑战性的、不受控制的现实环境中长期运行；*(II)通过关注实现这一飞跃所需的关键技术差距和严格的现场测试，使这些技术能够转移到实际应用中。**为了实现这些目标，有必要进行广泛的理论分析和使用地面和空中机器人的重要实验活动。上述两个目标可以进一步划分为未来五年要解决的几个短期目标：**为非受控环境中的操作设计特定任务的学习策略；*设计任务独立学习的方法；*为异质和同质机器人团队开发协作学习方案；并演示在不同机器人平台上长时间的成功学习。**申请人将与加拿大公司和组织合作，如MDA， CSA， Clearpath和PrecisionHawk。这些合作促进了在商用空中和地面机器人平台上对算法的广泛测试，并为申请人提供了最终用户的反馈。最终，根据申请人的研究结果，加拿大工业将有巨大的机会开发个人、企业和空间机器人解决方案，并将其目前的投资组合多样化到数十亿的应用领域，如遥感、自然资源监测和监视。