Chapter 40 Springer Handbook of Robotics, ©2008 Presented by:Shawn Kristek
1. Task Complexity 2. Distributed tasks 3. Difficult to build ultimate robot 4. Parallelism 5. Robustness through redundancy
Architectures Communication Variance Task Allocation Learning Applications
Architectures Communication Variance Task Allocation Learning Applications
o Centralized o Hierarchical o Decentralized o Hybrid
Centralized Single point of control Works best when controller oversees other robots X Vulnerable to single robot failure X Real-time difficulties due to communication requirements
Hierarchical Similar to military command More resistant to single robot failures X Vulnerable to upper-level single robot failure
Decentralized Most common Each robot’s actions based on localized data Robust to single robot failure X Global coherency difficult -Incorporated high-level goals difficult to revise Matarić
Hybrid Combinations of other architectures Advantages of levels of control and localized control -Robust to failures -Global coherency DIRA
Implementations The NERD Herd ALLIANCE DIRA
Implementations: The NERD Herd Matarić Decentralized Stigmergic Swarm robots Homogeneous: 20 identical robots Behavior-based Applications -Foraging & Coverage -Flocking & Formations The NERD Herd
Implementations: ALLIANCE Parker Decentralized Minimal explicit communication Heterogeneity possible Behavior-based - Uses motivations Applications -Box pushing & Cooperative Manipulation -Multitarget observation ALLIANCE
Implementations: DIRA DI stributed R obot A rchitecture Simmons Hybrid Explicit communication Heterogeneity possible Applications - Cooperative Manipulation DIRA
Architectures Communication Variance Task Allocation Learning Applications
o Stigmergy o Passive action recognition o Explicit
Stigmergy Sense through the world Simple No communication channels & protocols X Limited by robot’s perception Melhuish and Holland
Passive action recognition Communication through observation No limited bandwidth No fallible mechanism X Limited by robot’s perception X Difficult to analyze actions
Explicit Direct communication -Synchronize actions -Exchange information -Negotiate Directness Ease of acquiring knowledge of teammates X Noisy, limited-bandwidth channel
Architectures Communication Variance Task Allocation Learning Applications
o Swarm Robots – Homogeneous o Heterogeneous
Swarm Robotics Collective robotics Typically homogeneous Biologically inspired -Ants -Bees Stigmergic Redundant
Heterogeneous More realistic: - Heterogeneity may emerge in homogeneous systems Provides various capabilities Can reduce costs X Unavoidable Parker Grabowski
Architectures Communication Variance Task Allocation Learning Applications
o Taxonomy o Approaches
What is it? Efficiently assign tasks Team goal defined as set of tasks Each task can be subdivided Goal
Taxonomy Gerkey and Matarić Tasks -SR : Single-robot task -MR: Multirobot task Robots -ST: Single-task robot -MT: Multitask robot Allocation Optimization - IA: Instantaneous Assignment -TA: Time-extended Assignment SR-ST-IA
Approaches Behavior-Based Market-Based
Behavior-Based Decentralized architecture Avoids explicit communication Task Allocation -Current state -Teammate capabilities
Market-Based Negotiation/bidding based Greedily assigned to robot with highest utility Most focus on SR-ST-IA/TA Centralized or Hybrid architecture Explicit communication M+ architecture of Botelho and Alami -First for multirobot -Individual plans merged for team benefit
Architectures Communication Variance Task Allocation Learning Applications
Many difficulties -Exponential state spaces -Limited training time -Insufficient data -Uncertainty -Merging information Applied Applications -Multitarget observation -Box pushing -Multirobot soccer Techniques -Reinforcement -Parameter tuning -Particle swarm optimization
Architectures Communication Variance Task Allocation Learning Applications
o Foraging & Coverage o Flocking & Formations o Box Pushing & Cooperative Manipulation o Multitarget Observation o Traffic Control & Multirobot Path Planning o Soccer
Foraging & Coverage
Flocking & Formations Formation Control
Box Pushing & Cooperative Manipulation Kube
Multitarget Observation Spletzer and Taylor
Traffic Control & Multirobot Path Planning Bruce and Veloso
Soccer Bruce and Veloso RoboCup