Efficient Learning from Demonstration for adaptation of behaviour based on force feedback

基于力反馈的行为适应示范的高效学习

• 批准号: 2612195
• 金额: --
• 依托单位: University of Edinburgh
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2021
• 资助国家: 英国
• 起止时间: 2021 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-2612195
• 关键词: Efficient; Learning; Demonstration; adaptation; behaviour

Today, robots are used in various aspects of our lives, and there are many situations where humans and robots collaborate or work closely together. However, there are still many challenges to be overcome before robots can be used in a wider range of fields. Robots that perform everyday tasks, such as cooking, or that work with humans in rehabilitation or assisted tasks, still have problems with the effort required for learning, adaptability and safe operation around humans. Therefore, my project aims to develop a method for robots to efficiently learn motor skills involving force interaction from human demonstrations and to adapt their behaviour using force feedback from the objects they manipulate or the human partner they interact with. It also aims to develop a framework to enable the robot to explain its behaviour, failures and understanding of the task and environment so that it can operate safely around humans. From a practical point of view, efficient learning of tasks and improved adaptability and explainability will contribute to the wider use of collaborative robots, especially in areas where physical human-robot interaction is required, such as manufacturing, hospitality, nursing care and welfare. To achieve this goal, I will explore how robots can learn to obtain information about the objects they manipulate or the human partner they interact with and use it to adapt their behaviour according to the properties of the objects such as hardness and adhesiveness, and human capabilities such as joint flexibility and muscle strength, through Learning from Demonstrations. Learning from Demonstration is a fast, intuitive, and flexible way of transferring motor skills from humans to robots, however, teaching how to interact with the world and use force feedback to adapt behaviour can take a lot of time and effort for human demonstrators. Therefore, in my project, I will also develop an algorithm that allows efficient learning of motor skills involving force interaction by taking two approaches: improving the way of giving demonstrations and allowing to learn tasks in a way that is transferable across different tasks, objects, humans and tools so that they do not have to re-learn for each task, object, person or tool. Finally, since safety is extremely important, especially for robots working closely or together with humans, I will explore how robots can compute the uncertainty in their actions and their causes and explain the reasons for their actions and failures. In my first year, I will explore efficient ways of teaching robots to identify object properties using force interactions and verify the methods for adapting their behaviour based on identified object properties in daily manipulation tasks such as cutting and stirring. In my second year, I will investigate how robots can learn motor skills in one task and apply them to another task without re-learning. I will also work on developing a framework for robots to compute uncertainties in their behaviour and their causes and to explain the reasons for their actions and failures. In my third year, I will explore how these methods and frameworks can be applied to modelling of the physical characteristics and capabilities of individual human partners to enable smooth and safe human-robot collaboration.

今天，机器人被用于我们生活的各个方面，人类和机器人合作或紧密合作的情况很多。然而，在机器人应用于更广泛的领域之前，仍有许多挑战需要克服。执行日常任务的机器人，如烹饪，或与人类一起进行康复或辅助任务的机器人，在学习、适应性和人类周围的安全操作方面仍然存在问题。因此，我的项目旨在开发一种方法，让机器人有效地学习涉及人类演示的力交互的运动技能，并利用它们操纵的物体或与它们交互的人类伙伴的力反馈来调整它们的行为。它还旨在开发一个框架，使机器人能够解释其行为，故障以及对任务和环境的理解，以便它可以在人类周围安全操作。从实际的角度来看，对任务的有效学习以及改进的适应性和可解释性将有助于更广泛地使用协作机器人，特别是在需要人机交互的领域，如制造业、酒店业、护理和福利。为了实现这一目标，我将探索机器人如何学习获取关于它们操纵的物体或与它们互动的人类伙伴的信息，并利用它根据物体的属性（如硬度和粘附性）以及人类的能力（如关节灵活性和肌肉力量）来调整它们的行为。从演示中学习是一种快速，直观和灵活的方式，将运动技能从人类转移到机器人，然而，教如何与世界互动并使用力反馈来适应行为可能需要大量的时间和精力为人类演示。因此，在我的项目中，我还将开发一种算法，通过两种方法来有效地学习涉及力交互的运动技能：改进演示的方式，并允许以一种可在不同任务、物体、人和工具之间转移的方式学习任务，这样他们就不必为每个任务、物体、人或工具重新学习。最后，由于安全是极其重要的，特别是对于与人类密切合作或一起工作的机器人，我将探讨机器人如何计算其行为及其原因的不确定性，并解释其行为和失败的原因。在我的第一年，我将探索教机器人使用力相互作用识别物体属性的有效方法，并验证基于识别的物体属性在日常操作任务（如切割和搅拌）中调整其行为的方法。在我的第二年，我将研究机器人如何在一个任务中学习运动技能，并将它们应用到另一个任务中，而不需要重新学习。我还将致力于为机器人开发一个框架，以计算它们行为中的不确定性及其原因，并解释它们的行为和失败的原因。在我的第三年，我将探索如何将这些方法和框架应用于个体人类伙伴的物理特征和能力的建模，以实现顺利和安全的人机协作。