Improving robots with self-learning ability is one of the critical challenges for the researchers in the area of cognitive robotics and artificial general intelligence. This robot will decide when, where, and what to learn in a continuous visual environment by itself. Here we focus on the procedural knowledge learning, which is sequential and considered harder to understand compared with declarative knowledge in the cognitive system. Inspired by the architecture of the human brain which has integrated well different kinds of cognitive functions, a Brain-inspired Active Learning Architecture (BALA) is proposed for procedural knowledge understanding based on Baxter robot and human interaction. The BALA model contains four main parts: inspired by Primary Visual Pathway, a Convolutional Neural Network (CNN) is constructed for spatial information abstraction; inspired by the Hippocampus Pathway (especially the recurrent loops in CA3 sub-region), a Recurrent Neural Network (RNN) is built for sequential information processing related with procedural knowledge; inspired by the Prefrontal Cortex, a Knowledge Graph based on Bag Of Words (BOW) is constructed for declarative knowledge generation and association; inspired by the Basal Ganglia Pathway, we select Q matrix for Reinforcement Learning (RL). The CNN and RNN parts will be firstly pre-trained on ImageNet dataset and standard Youtube Video-Scene dataset respectively. Then, the RNN, Knowledge Graph, and Q matrix will be dynamically updated in the Baxter robot’s interactive learning procedure with human cooperators. The BALA could actively and incrementally recognize different kinds of procedural knowledge. In 22-type daily-life videos with procedure knowledge (e.g., opening the door, wiping the table, or taking the phone), the BALA model gets the best performance compared with standard CNN, RNN, RL, and other integrative methods. The BALA model is a small step on integrative intelligence interaction between the Baxter robot and human cooperator.

提高具有自学习能力的机器人是认知机器人和人工智能领域的研究人员面临的关键挑战之一。该机器人将自行决定在连续视觉环境中何时，何地以及要学习什么。在这里，我们重点关注过程知识学习，它是顺序的，与认知系统中的声明性知识相比，被认为更难理解。受集成了多种认知功能的人脑结构的启发，提出了一种基于脑的主动学习体系（BALA），用于基于Baxter机器人和人类交互作用的过程知识理解。BALA模型包含四个主要部分：受主要视觉通路的启发，构建了卷积神经网络（CNN）用于空间信息抽象；受海马通路（尤其是CA3子区域的循环回路）的启发，建立了循环神经网络（RNN），用于处理与过程知识相关的顺序信息；受前额叶皮层的启发，构建了基于词袋（BOW）的知识图，用于声明性知识的产生和关联。受基础神经节通路的启发，我们选择了强化学习（RL）的Q矩阵。CNN和RNN零件将首先分别在ImageNet数据集和标准的Youtube Video-Scene数据集上进行预训练。然后，将在百特机器人与人类合作者的交互式学习过程中动态更新RNN，知识图和Q矩阵。BALA可以积极地逐步认识到不同种类的程序知识。在具有程序知识（例如开门，擦桌子或拿电话）的22类日常生活视频中，与标准CNN，RNN，RL和其他集成方法相比，BALA模型具有最佳性能。BALA模型是百特机器人与人类合作者之间的集成智能交互的一小步。