The paper presents a new architecture for controlling mobile robots, emphasizing layers of control systems that allow the robot to operate at increasing levels of competence. Each layer is composed of asynchronous modules communicating over low-bandwidth channels, with each module being a simple computational machine. Higher-level layers can subsume the roles of lower levels by suppressing their outputs, while lower levels continue to function even as higher levels are added, ensuring robustness and flexibility. The system has been tested on a mobile robot navigating unconstrained laboratory areas and computer machine rooms, with plans to extend it to map office environments and perform simple tasks. The authors decompose the control problem into task-achieving behaviors rather than functional modules, leading to a different implementation strategy. The architecture is designed to handle multiple goals, sensors, robustness, and extensibility. The control system is built using finite state machines and asynchronous communication, with each module running independently and suppressing or inhibiting the outputs of other modules. The paper also includes a detailed specification language for the modules and their communication, and discusses the performance of the system in both simulated and physical environments. The key ideas are the decomposition of the control problem into behaviors, the incremental build and test of complex systems, the ability to perform useful parallel computation on a loosely coupled network of simple processors, and the absence of a central control module.The paper presents a new architecture for controlling mobile robots, emphasizing layers of control systems that allow the robot to operate at increasing levels of competence. Each layer is composed of asynchronous modules communicating over low-bandwidth channels, with each module being a simple computational machine. Higher-level layers can subsume the roles of lower levels by suppressing their outputs, while lower levels continue to function even as higher levels are added, ensuring robustness and flexibility. The system has been tested on a mobile robot navigating unconstrained laboratory areas and computer machine rooms, with plans to extend it to map office environments and perform simple tasks. The authors decompose the control problem into task-achieving behaviors rather than functional modules, leading to a different implementation strategy. The architecture is designed to handle multiple goals, sensors, robustness, and extensibility. The control system is built using finite state machines and asynchronous communication, with each module running independently and suppressing or inhibiting the outputs of other modules. The paper also includes a detailed specification language for the modules and their communication, and discusses the performance of the system in both simulated and physical environments. The key ideas are the decomposition of the control problem into behaviors, the incremental build and test of complex systems, the ability to perform useful parallel computation on a loosely coupled network of simple processors, and the absence of a central control module.