Autonomous Intelligent Machines & Systems

研究背景和潜在影响服务机器人可以彻底改变整个面向服务领域的工作方式，其潜在的应用包括在通常是远程的工业设施中进行自主维护[1,2]，如石油钻井平台，或在家庭环境中护理机器人[4,5,6]。为了有效地完成任务，机器人利用对环境的观察来构建和执行合适的计划。为了使其具有可扩展性，机器人必须能够泛化到它可能遇到的任何任务。虽然存在针对特定任务类别（例如，打开具有不同几何形状的抽屉）的通用操作技巧的方法，但当不同任务具有根本不同的完成条件并需要不同的操作策略（例如，按下按钮与拧紧阀门）时，这个问题变得更加具有挑战性。对于一个机器人系统来说，它具有重要的价值，它可以对以前看不见的任务进行最小的语义描述（例如“拧紧阀门”），自主理解并生成完成该任务所需的步骤序列，并相应地合成操作策略。目的和目标该项目将研究基于算法和学习的操作策略综合，使机器人能够以最少的人工输入推广到新的操作任务。这包括用自然语言或形式语言对任务的语义、概念描述进行综合。理想情况下，这些描述应该是任务成功的要素（例如，阀门是紧的），而不是如何实际完成任务；这是机器人要解决的问题。与EPSRC战略和研究领域保持一致本项目与EPSRC机器人1和机器人2研究领域直接相关。公司或合作者的参与我们计划与Alessandro Abate教授和牛津控制与验证小组（OXCAV）合作完成项目的关键部分，特别是操作策略的自动合成，以推广到新的操作任务。

Research Context and Potential ImpactService robots can revolutionize how entire service-oriented domains work, with potential applicationsincluding autonomous maintenance in - often remote - industrial facilities [1, 2], such as oil rigs [3], orcare robots in domestic environments [4, 5, 6].To complete tasks effectively, robots use observations from the environment to construct and executea suitable plan. For this to be scalable, the robot must be able to generalize to any task that it mayencounter. While methods exist for generalizing manipulation skills for a particular class of task [7] (forexample, opening drawers with different geometries), this problem becomes far more challenging whendifferent tasks have fundamentally different completion conditions and require different manipulationstrategies (for example, pressing a button versus tightening a valve). There would be significant valuein a robot system that can, given a minimal, semantic description of a previously unseen task (eg 'twistthe valve until tight'), autonomously understand and generate the sequence of steps needed to completethat task and synthesize a manipulation strategy accordingly.Aims and ObjectivesThe project will investigate algorithmic and learning-based synthesis of manipulation strategies that enable a robot to generalize to new manipulation tasks with minimal human input. This includes synthesisfrom semantic, conceptual descriptions of the task written in natural or formal language. Ideally, thesewould be descriptions of what constitutes task success (for example, the valve being tight) rather thanhow to actually accomplish the task; this is for the robot to figure out.Alignment to EPSRC Strategies and Research AreasThis project directly relates to the Robotics1 and AI2 EPSRC research areas.Involvement of Companies or CollaboratorsWe plan to collaborate with Prof. Alessandro Abate and the Oxford Control and Verification group(OXCAV) for key parts of the project, in particular the automatic synthesis of manipulation strategiesto generalize to new manipulation tasks.