The Control Challenge in Mobile Robotic Systems
A mobile robot is much more than a machine that moves. It integrates three essential systems: mechanical, electronic, and cognitive. The latter, the processing and decision-making system, is the core of the control architecture. Designing such an architecture involves defining how components relate, their functions, and how they communicate. In dynamic environments, the system must process continuous data, discrete events, and noisy sensory information, all in real time. The architecture is not just a technical blueprint: it is the distributed brain of the robot.
Shakey: The Pioneering Feat of AI Planning
Between 1966 and 1972, SRI International developed Shakey, the first mobile robot capable of reasoning about its own actions. Funded by DARPA, Shakey marked a before and after in intelligent robotics. Its top-down hierarchical architecture was based on an axiomatic model of the world, which fed the STRIPS planning system. This approach allowed complex orders to be broken down into basic actions, without the need for step-by-step programming. Shakey did not just execute: it understood, decided, and acted.
The Deliberative Paradigm (SPA): The Initial Domain of AI
Inspired by symbolic AI, the deliberative paradigm dominated early robotic control designs. Known as SPA (Sense – Plan – Act), this model required the robot to first sense its environment, then build a detailed world model, plan its action, and finally execute it. Examples like NASREM (NASA/NBS) replicated this hierarchical structure. Although powerful in controlled environments, SPA showed limitations in dynamic and unpredictable scenarios.
Challenges and the Need for Reactivity
The rigidity of the deliberative model soon became problematic. Symbolic planning required time and precision, making it difficult to respond quickly to unexpected changes. In dynamic environments, plans could become obsolete before they were executed. This operational slowness led to a decade of obsession with world modeling, but it also opened the door to new, more agile, and adaptive forms of control.
The Reactive Architecture (SA): Prioritizing Agility
Rodney Brooks revolutionized the field with the Subsumption Architecture, the basis of the reactive paradigm. This approach eliminated centralized planning and was inspired by animal behavior: direct sensing and action. Behavior layers operated asynchronously, and higher layers could inhibit lower ones based on priority. Thus, obstacle avoidance could suppress wandering. The result: robust, fast, and adaptive systems, where intelligence emerged from the interaction between simple modules.
Hybrid Architectures: The Balance Between Planning and Reflex
During the nineties, hybrid architectures emerged that combined the best of both worlds. They were structured in three layers: a reactive one for immediate control, an intermediate one for sequencing behaviors, and a deliberative one for complex planning. This design allowed the robot to respond quickly to stimuli while planning in the background. Examples like AuRA, 3T, and Saphira showed that hierarchical integration could be efficient and scalable.
The Evolution Towards Multi-Agent Systems (MAS)
The multi-agent paradigm represents a conceptual evolution: distributing intelligence among multiple autonomous agents. Originating from Distributed Artificial Intelligence (DAI), MAS allow each module to have its own goals, capabilities, and decisions. This modularity facilitates scalability and the integration of new functions without redesigning the entire system. Instead of a monolithic architecture, a distributed cognitive ecosystem is built.
AGC-MAS: The Generic Multi-Agent Control Architecture
AGC-MAS proposes a higher abstraction for designing robotic systems as networks of intelligent agents. Independent of the underlying middleware, this architecture allows for modeling teleoperated or autonomous robots, whether deliberative, reactive, or hybrid. Its modular approach facilitates component reuse, the insertion of new agents, and adaptation to different platforms. AGC-MAS does not impose a rigid structure but offers a flexible framework for distributed cognitive design.
Hierarchical Structure and Topology of AGC-MAS
The AGC-MAS architecture is organized into four logical levels: System, Subsystem, Node, and Agent. The System groups one or more Subsystems (like the mobile platform or the manipulator arm). Each Subsystem contains Nodes that offer specific services (vision, navigation, etc.). At the base, Agents execute concrete functions and can deliberate on the use of computational resources. This hierarchy allows for a clear and scalable organization of robotic control.
Functional Classification and Roles of Agents
AGC-MAS classifies agents according to their function. Exteroceptive Agents (ID 00–09) capture physical variables like position or distance. Reactive Agents (ID 10–19) respond to immediate stimuli, such as avoiding obstacles. Deliberative Agents (ID 40–49) model the environment and plan missions. The Mobile_Robot Agent (ID 30) represents the physical plant. This classification allows for designing generic and adaptable systems, where each agent fulfills a specific role within the ecosystem.
Semantic Language and Timestamp
Communication between agents is done through structured messages in Reports, Requests, and Commands. Each message includes a header with origin, destination, and a global Timestamp. This timestamp allows agents to evaluate the currency of data and decide where to execute their functions. Inspired by standards like JAUS, this approach promotes interoperability and code reuse, facilitating the development of robust and modular systems.
Standardization and Interoperability in Modern Robotics
Despite advances, robotics still faces a critical challenge: the lack of universal technical standards. Middleware like Player, SmartSoft, or RT-Middleware, and simulators like Stage or Gazebo, offer powerful tools, but without a common interface, code exchange remains complex. Architectures like AGC-MAS, with semantic support and explicit modularity, are key to advancing towards plug-and-play robotics, where components can be integrated without friction.