基于认知学习的智能机器人控制系统关键问题的研究

Research in cognitive robotics is concerned with endowing robots with the capability of cognitive learning that enable them to perceive-learn-act in dynamically changing environments. In this proposal three key problems, namely the feature self-organizing of the sensory dada of the dynamically changing environments, the perception-action mapping and optimizing for the behaviors required, and the adaptive online learning control of robots in real time, are addressed based on the notion of long and short term memory, dual coding, and visual attention and reward in cognitive science. Specifically, we address the self-organizing module, spatial salience detection, attention selection and reward, and the invariance of the mental rotation transformation of the visual image from low level sensor data; the self-growing and self-pruning hierarchical neural network models in Unsupervised-Supervised-Reinforcement (UL-SL-RL) learning paradigms, the adaptively online tuning and Q-learning to the behavior controller parameters, and the stability analyses in terms of the experimental evaluation on the real robot platforms, finally resulting in a perceive-learn-act cognitive control system for autonomous robots. The ultimate goal of the project proposed is aiming at the learning mechanisms in cognitive robotics that enable robots to learn new behaviors and to tune the existing behaviors to facilitate the dynamically changing environments, consequently increasing the cognitive learning ability of the robots, improving the universality and generalization of the learning system, and simplifying the controller design.

认知机器人研究的是赋予机器人认知学习能力使之能在动态变化的环境下自主实现感知-学习-动作的认知过程。本项目应用认知科学中长短期记忆模型、双向编码机理、视觉注意与激励等机制，对机器人在动态环境下感知数据的特征自组织模型、所需行为从感知到动作的映射与优化、和机器人实时在线自适应自学习控制三个关键问题开展研究。具体地，我们重点研究视觉图像的空间显著特征检测和自组织模式、视觉注意力选择和旋转变换保持机制，"非监督-监督-激励"集成学习模式下机器人状态-动作映射的自生长自消减多层次神经网络模型，行为控制器参数的实时自适应学习和在线策略梯度Q-学习优化，以及基于机器人实验研究的学习系统稳定性评价方法，最终形成一个感知-学习-动作循环的自主机器人学习系统。本项目旨在揭示机器人的学习控制机理、使之具备学习新行为和优化现有行为的能力，从而提高机器人的认知学习水平、增强学习系统的普适性、和简化控制器的设计。

本项目应用认知科学中长短期记忆模型、双向编码机理、视觉注意与激励等机制，对机器人在动态环境下感知数据的特征自组织模型、所需行为从感知到动作的映射与优化、和机器人实时在线自适应自学习控制三个关键问题开展研究。具体地，我们研究了视觉图像的空间显著特征检测和自组织模式、视觉注意力TD-BU选择和旋转变换保持机制，“非监督-监督-激励”集成学习模式下机器人状态-动作映射的自生长自消减多层次神经网络模型，行为控制器参数的实时自适应学习和在线策略梯度Q-学习优化，以及基于机器人实验研究的学习系统稳定性评价方法，最终形成了一个感知-学习-动作循环的自主机器人学习系统。通过本项目的研究，我们获得了两项国家发明专利授权，申请了4项国家发明专利；发表了相关论文5篇，培养了硕士研究生8名。