The Subsumption Architecture

Traditional Robotics

• Deliberative control

• Knowledge-based approach

• Maintain an accurate world model internally

• Path planning

• Sequential sense->model->plan->action organization

• Poor performance in real-time complex environments

• Complex behavior results from complex internal algorithms

• Traditional task decomposition:

  

Behavior-Based Robotics

• Reactive control

• "World is its own model"

• No explicit internal representations

• Direct sensory-motor couplings

• Distributed asynchronous organization

• Good performance in real-time complex environments

• Complex behavior results from interactions among simple internal processes and environment

• "Ant wandering on a beach" example

• Behiavor-based task decomposition:

  

Subsumption Architecture

• Layers of control

• Incremental, evolutionary approach

• Inter-layer suppression and inhibition

  

• Modules and augmented finite state machines

  

• Example (Brooks, 1986):

  • Robot explores its environment, seeking out "interesting" places while avoiding obstacles
  • Layer 0: avoid obstacles
  • Layer 1: wander around aimlessly
  • Layer 2: seek out interesting places

  

Some Subsumption-Based Robots

• Genghis and Hannibal

         

• Cog