集成情景-程序性记忆认知特性的机器人经验学习与技能获取

Advanced theories can be learned from human cognitive process of producing new behavior through recalling experiences by memory. Episodic memory is an important intermediary between experience and behavioral skills, while procedural memory is the representation form of behavioral skills. The interaction between these two memories creates a new way for robot learning new skills. This project proposes a skill learning and acquisition method through integrating episodic memory and procedural memory, in order to improve the robotic adaptability to environments and tasks on the cognitive level. An episodic memory and procedural memory model is set up to express the episodic experiences and behavioral skills that robot experienced in space-time environments. Based on the self-organizing learning of hierarchical demonstration strategy memory, a universal learning paradigm to solve the interaction problem between robot and environment is built to form the comprehensive statement of episodic experiences and behavioral skills. Robotic cognitive behavior prediction and planning method based on episodic memory is studied. Skill acquisition strategy is proposed based on episodic experience -guided reinforcement learning. So that the robot can realize the autonomous planning, prediction and reasoning, and skills acquisition at a higher cognitive level. The project proposed new ideas and approaches on robotics experience and skill modeling, strategy self-organizing learning, real-time planning and skills acquisition from the perspective of human memory, and provided theoretical and technical support for the robotic skills acquisition inspired by human memory.

人类通过记忆回忆经验产生新行为的认知过程具有值得学习的先进理论。情景记忆是经验和行为技能间的重要中介，程序性记忆作为行为技能的表征形式，两者的交互为机器人学会新技能提供了新思路。本项目创新提出集成情景-程序性记忆研究机器人技能学习和获取方法，在认知层面上提高机器人对环境和任务的适应性。提出一种情景记忆与程序性记忆模型表述时空中机器人经历的情景经验和行为技能，基于分层示教策略记忆自组织学习，构建通用的解决机器人与环境间互动问题的学习范式，形成机器人对环境经验和行为技能的综合表述。研究基于情景记忆的机器人认知行为预测与规划方法，提出情景经验引导下基于强化学习的技能获取策略，使机器人在较高的认知水平上实现自主规划、预测与推理以及技能获取。项目以类人记忆角度对机器人经验和技能建模、策略自组织学习、实时规划和技能获取等科学问题提出新思想和方法，为基于类人记忆启发研究机器人技能获取提供理论与技术支撑。

人类通过情景记忆回忆过去经验，基于事件再配置产生程序性行为解决方案的认知过程具有值得学习的先进理论，此认知过程为机器人技能学习过程复用经验、快速学习提供了新途径。本项目创新提出基于情景-程序性记忆研究机器人技能学习与经验复用，在认知层面上提高机器人对复杂环境和任务的适应性。受人类记忆的认知机理启发，构建机器人情景记忆模型，实现机器人记忆的存储、检索与更新，基于对象交互进行情景-程序性记忆建模，指导机器人操作任务技能学习；提出一种示教策略模仿学习架构，可在无人为干预的情况下从示教中学习策略；基于以往的经验知识，提出一种情景引导下基于元学习的机器人技能学习方法，实现少样本情况下机器人对新任务的高效学习；结合情景记忆的认知特性，提出了情景引导的两相强化学习方法，提高经验的复用效率，减少学习新技能花费的时间。实验结果表明，基于情景记忆的学习方法相比无情景记忆学习方法平均奖励提升了超30%，任务学习平均成功率提升了超20%，可一定程度上实现对多种不同任务的泛化学习。项目以情景-程序性记忆角度对机器人操作经验快速学习和任务泛化科学问题提出新思想和方法，为基于情景-程序性记忆研究机器人认知行为提供理论与技术支撑。

内容获取失败，请点击重试