Behavior Based Systems

1
Behavior Based Systems

• Ramin Mehran
• Digital Control Laboratory
• K.N.Toosi U of Tech.
• Supervisors
• Professor Caro Lucas
• Dr. Alireza Fatehi

2
Contents

• Where BBS Stands?
• Functional/Task Decomposition
• Robotic Problem BBS test bed
• Reactive/BBS/Hybrid Arch.
• Subsumption Arch.
• Expressing Behaviors
• Behavior Coordination/Arbitration
• Learning/Robustness/Stability/Optimality
• BBS and Context-Based Systems
• Conclusion

3
Inter-disciplinary View

• System Theory
• Cybernetics, Control Theory
• Artificial Intelligence
• Intelligent Control
• New AI
• Behavior based Systems (Control)

4
What is a control problem?
Concept of Control
Action
Actuator output
Controller
set point
5
Functional Decomposition
Recognition
Action
Perception
Planning
Map building
sensors
actuators
6
Functional Decomposition
Control System
Recognition
Action
Perception
Planning
Map building
sensors
7
Robotics Increasing Complexity

• Dynamic and Nondeterministic Environment
• Nonholomic
• Conf. space smaller than Cont. space
• Sensors
• Similar to Real-Life Problems
• Failed Approaches

8
Brooks Critiques

• Engineering
• Robustness, Extendibility, Multiple goal, etc.
• Biological Inspirations
• Subtracts are used to build more complex
  capabilities

9
Brooks Critiques (cont.)

• Philosophical Inspirations
• Learning
• Unpredictability

Media Lab MIT - Leonard
10
Brooksian Manifesto
Intelligence is in the Eye of the Observer
11
Call for change The Architecture
Recognition
Action
Perception
Planning
Map building
12
Two Orthogonal Flows
Planning
Planning
World Model
World Model
Sensor
Motor
Sensor/Motor Control
Sensor/Motor Control
Sensor
Motor
13
Different Architectures

• Planer based control
• Moravec
• Reactive control
• Connel
• Hybrid control
• Arkin
• Behavior-based control
• Brooks

14
Behavior Based Properties

• No Global Representation
• e.g. No global map
• Are feedback controllers
• FSM, Fuzzy, PID, etc.
• Achieve specific tasks/goals
• (e.g., avoid-others, find-friend, go-home)
• Executed in parallel/concurrently
• Can store state and be used to construct world
  models
• (local representation)
• Behaviors can directly connect sensors and
  effectors

15
Subsumption Architecture

• First BBS
• Hierarchical
• Levels of competence
• Incremental
• Extendable
• Starting from most vital task

16
Structure of Modules in SSA
Inhibitor
Inputs
Outputs
Reset
Suppresor
17
SSA Example
Brook 1986
18
Hybrid Architecture
Hybrid Control!
Planner
Reactive / Behavior-Based
19
Expressing Behaviors

• Finite State Machine (FSM)
• Stimulus Response Diagrams
• Schema
• Fuzzy
• Potential Fields

20
Arbitration Which action has Control

• Subsumption has internal arbitration
• Inhabitation and suppression
• Hybrid Arch. Needs Beavior Arbitration
• Fuzzy Behavior Arbitration

21
Behavior Coordination

• Competitive
• Coordinative
• Combined
• Context-Dependent Blending

22
Mathematical Modeling

• Lack of Strict Modeling
• Poor Nonlinear Dynamic Modeling
• Stochastic Modeling for Learning
• FSM

23
Learning

• Reinforcement Learning
• Imitative Learning
• Learning Hierarchy, Behaviors, Sensor Fusion
• Credit Assignment Problem
• Evolutionary Algorithms

24
Optimality/Robustness/Stability

• Robust
• Failure in each part eliminates a task, not a
  full collapse
• Optimality measure as ave. reward
• Behavior Stability analysis
• No global stability analysis

25
BBS in Multi-Agent Systems

• Planner-based Arch. Fails for exponential growth
  of state space
• Uncertain and Unobservable
• Classical planning is intractable
• BBS uses local less complex strategies

26
BBS and Context-Based Systems

• Context is in the eye of the observer?!
• Hybrids are OK with context
• Pure BBS hard to show context transitions
• Creating new context?

27
Lit. of Context-Based BBS

• Arkins Case-based Schema

28
Lit. of Context-Based BBS (cont.)

• Bonarinis Fuzzy Brain

29
Lit. of Context-Based BBS (cont.)

• Saffiotis Context- based Behavior Blending

30
Conclusion

• When to use BBS and When to avoid it?
• Does it do real time?
• Do we know the model?
• How uncertain is the environment and sensors?
• When you can use a simple PID, use it!

31
Conclution (cont.)

• Pros
• Extendibility, Incremental, Real-world
  applicability, Robustness, Emergent, Modularity
• Most Real-world working robots are BBS!
• First 6 legged robot was Brooks!
• Cons
• No global representation, unclear design method,
  Stability, Optimality, Not explicit mathematical
  model

32
Thank you!
33
Moravecs Perspective
34
Potential Fields Expression
35
Schema Expression

• Low and High Gain