This research compares and analyses the Contributions and limitations of recent space robot control architecture and task-oriented multi-robot control architecture, presents a control architecture formally described by multi-agent system theory. The control architecture automatically change the relationships between agents according to the task description from user. User can interact with different agent to get feedbacks and make decisions of different resolution. Planning problem is distributed in several agents, so the problem solving will be quickly to meet the requirement of fault-tolerance. Detailed descriptions are as follows: (1) Firstly, it abstracts robots, the distributed closed-loop of sensor-actuator, software unit as agents; based on the abstraction, it then uses social institution of multi-agent system theory to design a task-oriented space robot control architecture; it divides social institution into two parts: a hierarchical organization model describes the task-oriented relationships between agents in the architecture, within the organization model interactions between agents are limited in a small scope to adapt to the application background of limited communication bandwidth and large-time delay; social norms compose of knowledge interaction norm which defines the data format and task interaction norm which defines the data content. (2) Secondly, it achieves knowledge interaction norm with a knowledge servicer and a interface criterion: the knowledge servicer manages a shared ontology which defines the vocabulary and constraints in communication, and provides consistency reasoning service to decide the knowledge representation of an agent is consistent with the shared ontology; the interface criterion uses KIF(Knowledge Interchange Format) as intermediate language to translate formats between different languages, and uses OKBC(Open Knowledge Base Connectivity) as protocol to enable agents access the knowledge servicer. (3) Thirdly, it achieves task interaction norm with a cognitive model based on SOAR(State, Operator and Result) as the theoretical basis for agent design, the cognitive model divides the internal structure of agent into five components: perception, working memory, long-term knowledge memory, cognitive processing and behavior output; it presents a logic system for decision-making, through the proofs for a target formula, a set of production rules will be generated for checking and analyzing the design of an agent; it then develops a platform by which developer can online debug agents for simulation and verification of control architecture. (4) It presents a online planning unit which is called by agent to describe and plan complex task that cannot be represented by target formula. The online planning unit describes task, domain model, resource and capability by NDDL(New Domain Description Language) and uses PSEngine which is provided by EUROPA(Extensible Universal Remote Operations Planning Architecture) as problem solving engine. If there is a plan in the plan database with no flaw, then agent compares the plan with temporal database which maintains a execution history to set current target, else then agent feed task compromise analysis to parent node. (5)Finally, it applies the control architecture to a space robot system and emphatically analyzes the advantage that planning problems are restrict within single agent; it indicates that problem solving will be more quickly introducing some heuristics strategy; it shows unique features and contributions by comparing itself with others task-oriented multi-robot control architectures.