Introduction to Robotics CSCI 445 Laurent Itti Group Robotics Introduction to Robotics L. Itti & M. J. Mataric 1
Today s Lecture Outline Defining group behavior Why group behavior is useful Why group behavior is hard How group behavior can be controlled Approaches to group behavior Examples Introduction to Robotics L. Itti & M. J. Mataric 2
Robot team escorting a moving target Introduction to Robotics L. Itti & M. J. Mataric 3
Team manipulating an object Introduction to Robotics L. Itti & M. J. Mataric 4
Terminology Various rather interchangeable terms are used in this area: Group behavior / robotics Collective behavior / robotics Cooperative behavior / robotics Swarm robotics Multi-robot systems Introduction to Robotics L. Itti & M. J. Mataric 5
Benefits of Group Solutions Improved system performance usually in terms of speed of completion Improved task enablement Tasks which are not possible with single robot Distributed sensing Beyond the range (traditional vs. distributed) Distributed action at a distance Action at different location Fault tolerance through redundancy Introduction to Robotics L. Itti & M. J. Mataric 6
Distributed Sensing Robot 1 Robot 2 Robot 3 Introduction to Robotics L. Itti & M. J. Mataric 7
Distributed Sensing: Limited View Robot 1 Robot 2 Robot 3 Introduction to Robotics L. Itti & M. J. Mataric 8
Adding Communication Robot 1 Robot 2 Robot 3 Introduction to Robotics L. Itti & M. J. Mataric 9
Building a Global Map Robot 1 Robot 2 Robot 3 Introduction to Robotics L. Itti & M. J. Mataric 10
Merging Sensor Data Noisy sensors: 1D Example Introduction to Robotics L. Itti & M. J. Mataric 11
Merging Sensor Data 1D Example: Two sensors (on different robots) Introduction to Robotics L. Itti & M. J. Mataric 12
Result of Two Sensors Error in observation is reduced Introduction to Robotics L. Itti & M. J. Mataric 13
Negatives of Group Solutions Interference among robots Communication cost and robustness Equipment Noise Computational processing Uncertainty about other robots intentions Based on limited knowledge Overall system cost System complexity Introduction to Robotics L. Itti & M. J. Mataric 14
Types of Collective Systems Merely Coexisting: do not even recognize each other, merely as obstacles + no need for coordination - uncontrolled interference Loosely Coupled: sense each other and may interact, but do not depend on each other + Robust - Difficult to coordinate precise task Introduction to Robotics L. Itti & M. J. Mataric 15
Types of Collective Systems Tightly Coupled: cooperate on a precise task, usually by using communication, turn-taking, and other means of tight coordination - Depend on each other Introduction to Robotics L. Itti & M. J. Mataric 16
Example Domains Mere coexistence foraging Loosely coupled foraging collection distributed mapping Tightly coupled formations moving objects Introduction to Robotics L. Itti & M. J. Mataric 17
Competitive Domains Besides cooperation there is also competition Game scenarios are a good challenge for developing group robotics Real world scenarios have competitive elements (robots are always competing for space => interference) Introduction to Robotics L. Itti & M. J. Mataric 18
Interference Physical interference competition for physical resources, like space Task interference competition for task resources, like objects competition for winning resources, like goals, pieces, etc. Introduction to Robotics L. Itti & M. J. Mataric 19
Communication Provides synchronization of action Enables information exchange Allows for negotiations Is NOT essential for cooperation Louder is not necessarily better Introduction to Robotics L. Itti & M. J. Mataric 20
Control Approaches How can we control a group of robots? Two basic options exist: centralized control distributed control Various combinations of the two in a hierarchical form Introduction to Robotics L. Itti & M. J. Mataric 21
Centralized Control A single, centralized controller Input: information about all of the robots Output: the actions for all of them. Problems: requires a lot of information requires global communication slow to plan for many agents (global state space is huge) depends on the centralized controller Introduction to Robotics L. Itti & M. J. Mataric 22
Centralized Control Problems: creates a bottleneck not scalable very slow not robust Advantage: allows the computation of optimal solutions (at least in theory) at the group level Introduction to Robotics L. Itti & M. J. Mataric 23
Distributed Control Each robot uses its own controller to decide what to do. Advantages : no information needs to be gathered communication can be minimized or avoided (no bottle-neck) robots or sub-groups can fail group size can change dynamically scales well with increased group size individuals can adapt and improve Introduction to Robotics L. Itti & M. J. Mataric 24
Distributed Control Disadvantages The desired group-level collective behavior must be produced in a decentralized, non-planned fashion from the interactions of the individuals Designing individual/local behaviors that result in the desired group/global behavior is a VERY hard problem Introduction to Robotics L. Itti & M. J. Mataric 25
Group Control Architecture There are only 4 types of control arch s Deliberative Reactive Hybrid Behavior based Those are suitable for different types of group behaviors (how?) Introduction to Robotics L. Itti & M. J. Mataric 26
Deliberative Group Control Deliberative systems => centralized The single controller (on or off a robot): gathers the sensory data uses it all to make a plan for all robots sends the plan to each robot, and each executes it Introduction to Robotics L. Itti & M. J. Mataric 27
Hybrid Group Control Well suited for centralized control and can be used in distributed control: Centralized: Deliberative layer: perform the SPA loop Reactive layer: individual robots monitor their sensors, and update the planner with any changes Distributed: each robot runs its own hybrid controller needs info on all others to plan; synchronizing the plans is hard Introduction to Robotics L. Itti & M. J. Mataric 28
Reactive Group Control Well suited for the distributed approach each robot executes its own controller, can communicate and cooperate with others as needed the group-level behavior emerges from the interaction of the individuals Introduction to Robotics L. Itti & M. J. Mataric 29
Behavior-Based Group Control Well suited for the distributed approach: each robot behaves according to its own, local behavior-based controller each robot can also learn over time and display adaptive behavior the group-level behavior emerges from the interaction of the individuals the group-level behavior can also be improved and optimized Introduction to Robotics L. Itti & M. J. Mataric 30
Hierarchies in Groups Hierarchical approaches can be implemented with any of the controllers Fixed hierarchies: can be generated by a planner within a deliberative or hybrid system Goal keeper in RoboCup Dynamic, changing, adaptive hierarchies: can be formed by behavior-based systems Introduction to Robotics L. Itti & M. J. Mataric 31
Challenges Controlling groups of robots is even more difficult than controlling one robot, because: the environment is inherently dynamic there are more interactions to consider there is more uncertainty in the system Introduction to Robotics L. Itti & M. J. Mataric 32
Prototypical Group Tasks Foraging Distributed mapping/exploration Grazing/coverage Formations Object transport Sensor-actuator networks Robot soccer Introduction to Robotics L. Itti & M. J. Mataric 33
Why Foraging? Foraging is a prototype for useful (loosely coupled) tasks: locating and disabling/marking land mines distributed mapping of the area collectively distributing objects (markers, cables, seeds, etc.) collective reconnaissance collective surveillance and many more... Introduction to Robotics L. Itti & M. J. Mataric 34
Foraging using pheromones From HRL labs Introduction to Robotics L. Itti & M. J. Mataric 35
Why Box Pushing? Box pushing is a prototype for useful (tightly coupled) tasks: construction/destruction help in disaster scenarios moving wounded Challenges: Arbitrary object geometry Arbitrary numbers of robots Arbitrary initial configuration Homogeneous or heterogeneous teams Introduction to Robotics L. Itti & M. J. Mataric 36
Box pushing example Mataric lab @ USC Introduction to Robotics L. Itti & M. J. Mataric 37
Characteristics Reliability: The probability that a system acts correctly in the given situation. Organization: Jack of all trades master of none Communication: Information content: very small for animals Mode: method of communication Spatial distribution Small group vs. large group Overlapping vs. non-overlapping Congregation: How to stay together? Performance Speedup Energy used Introduction to Robotics L. Itti & M. J. Mataric 38
Taxonomy Team size Communication range Communication topology: Broadcast (one to many) Addressed (one to a group) Tree (one to children) Graph (one to a fixed set) Communication bandwidth Introduction to Robotics L. Itti & M. J. Mataric 39
Taxonomy Team reconfigurability Subclass static Communication coordinated (the robots in contact can reconfigure) Dynamic Team unit processing ability Team composition Homogeneous Heterogeneous Introduction to Robotics L. Itti & M. J. Mataric 40
Example: Nerd Herd Nerd Herd: a collection of 20 coordinated small wheeled robots (Mataric 1994) Basis behaviors: homing, aggregation, dispersion, following, safe wandering Organized in Subsumption style Complex aggregate behaviors: flocking, surrounding, herding, docking Complex behaviors result from combinations or sequences of basis set Introduction to Robotics L. Itti & M. J. Mataric 41
Nerd Herd Introduction to Robotics L. Itti & M. J. Mataric 42
Multi-robot Topics Communication Kin recognition Task allocation Introduction to Robotics L. Itti & M. J. Mataric 43
Communication Communication: Enables synchronization of behaviors Enables information sharing & exchange Enables negotiations Communication not necessary or essential for cooperation Louder is not necessarily better Introduction to Robotics L. Itti & M. J. Mataric 44
Communication Cost Communication is not free Hardware overhead Software overhead For any given robot task, it is necessary to decide: whether communication is needed at all what the range should be what the information content should be what performance level can be expected Introduction to Robotics L. Itti & M. J. Mataric 45
What to Communicate? State (e.g., I have the object) Goal (e.g., go this way, follow me) Intentions (e.g., I m trying to find the object, I m trying to pass you the ball) Representation (e.g., maps of the environment, knowledge about the environment, task, self, or others) Introduction to Robotics L. Itti & M. J. Mataric 46
Example: Foraging What could be communicated: nothing: by observation only (implicit or stigmergic communication) the location of the object the description of object the direction to go in locations/directions to avoid (due to interference, obstacles, danger, etc.)... Introduction to Robotics L. Itti & M. J. Mataric 47
Introduction to Robotics L. Itti & M. J. Mataric 48
Stigmergy Stigmergy is communication through sensing the effects of others in the environment (instead of using direct messages) Examples: ant trails grazing patterns piling up pucks/ant hills This powerful mechanism is common in nature and can be used cleverly Introduction to Robotics L. Itti & M. J. Mataric 49
Stigmergy Example Using MASyV (Multi-Agent System Visualization) http://masyv.sourceforge.net Introduction to Robotics L. Itti & M. J. Mataric 50
Kin Recognition Kin recognition is the ability to recognize others like me In nature, it usually refers to the members of the immediate family (shared genetic material); can be used for sharing of food, signaling, altruism In robotics, it refers to recognizing other robots (and other team-members) as different from everything else in the environment Introduction to Robotics L. Itti & M. J. Mataric 51
Kin Recognition Importance Without kin recognition, the types of cooperation that can be achieved are greatly diminished (mere coexistence) Kin recognition does not necessarily involve recognizing the identities of others, but if those are provided, more sophisticated cooperation is possible (dominance hierarchies, alliances, etc.) Ubiquitous in nature; methods in robotics: audio beacons visual beacons (active/passive) laser beacons Introduction to Robotics L. Itti & M. J. Mataric 52
Kin recognition methods Laser-visual beacon Introduction to Robotics L. Itti & M. J. Mataric 53
Applications Distributed sensing + coordinated movement result in: convoying (transportation: Follow the leader) landmine detection: Needs multiple sensors in different locations reconnaissance & surveillance blanket coverage: Watch for intruders at a taken station Introduction to Robotics L. Itti & M. J. Mataric 54
Applications Distributed sensing + coordinated movement result in: barrier coverage: A barrier of robots to detect crossing sweep coverage : Sweep for enemy map making : Introduction to Robotics L. Itti & M. J. Mataric 55
Multi-robot task allocation Task allocation: Who-does-what problem Given: m tasks and n robots utility for each robot/task pairing Find: optimal assignment of robots to tasks Assuming: single task per robot single robot per task Introduction to Robotics L. Itti & M. J. Mataric 56
MRTA Definitions: Utility = benefit - cost (robot-task pair) Total utility = sum of individual utilities Utility matrix: rows = robots columns = tasks elements = utility Optimal solution: maximizes total utility Introduction to Robotics L. Itti & M. J. Mataric 57
MRTA Tasks Robots Utilities / costs Introduction to Robotics L. Itti & M. J. Mataric 58
Greedy assignment Solution: greedy assignment First-come-first-served Pros: fast, simple Cons: formally sub-optimal Introduction to Robotics L. Itti & M. J. Mataric 59
Hypothesize and test Solution: hypothesize and test Simple-minded search Pros: simple, optimal Cons: very, very slow Introduction to Robotics L. Itti & M. J. Mataric 60
Hungarian method Solution: Hungarian method Exploit linear-algebraic structure Solves optimal task assignment (who does what matrix) in polynomial time Pros: optimal, fast Cons: complex, centralized Introduction to Robotics L. Itti & M. J. Mataric 61
Distributed Auctions Solution: iterated auction Robots bid on tasks Pros: optimal in limit, simple, distributed Cons: utility difficult to assign properly Introduction to Robotics L. Itti & M. J. Mataric 62
Other MTRA problems Assumptions in the discussion so far: single task per robot single robot per task linear utility combination static problem (offline) Real problems tend to violate assumptions: multiple robots per task (e.g., carrying) interdependent utility (e.g., fire-fighting) dynamic problem (online) No optimal solutions? Introduction to Robotics L. Itti & M. J. Mataric 63
Reading MM Ch 20 Introduction to Robotics L. Itti & M. J. Mataric 64