1. ALLIANCE: An Architecture for Fault Tolerant Multirobot Cooperation L. E. Parker, 1998 Presented by Guoshi Li April 25th, 2005

2. Presentation Outline Introduction Background Alliance Results Conclusion

3. Introduction For smaller-scale applications, the single robot approach is often feasible A large number of the human solutions to these real world applications of interest employ the use of multiple humans supporting and complementing each other The use of robot teams for automated solutions to some real applications is feasible and necessary

4. Introduction Advantages to using a distributed mobile robot system  Reduce the total cost of the system  Increase the robustness of the system by taking advantage of the parallelism and redundancy of multiple robots  Accomplish a mission which requires the use of multiple robots working simultaneously on different aspects with time constraints

5. Introduction Challenges of the use of multiple robots  May actually increase the complexity of an automated solution  Achieving coherence  Determining how to decompose and allocate the problem among a group of robots  Determining how to enable the robots to interact

6. Introduction Fault tolerance  The ability of the robot team to respond to individual robot failures of failures in communication Adaptivity  The ability of the robot team to change its behavior over time in response to a dynamic environment, changes in the team mission, or changes in the team capabilities or composition

7. Background Cooperative mobile robotics  Swarm-type cooperation  “Intentional” cooperation Swarm-type cooperation  Deals with large numbers of homogeneous robots  Useful for nontime-critical applications  Globally interesting behavior can emerge as a result of the local interactions of the robots  A key research issue: determining the proper design of the local control that will allow the collection of robots to solve a given problem

8. Background “Intentional” cooperation  Deals with a limited number of typically heterogeneous robots  Has to address some sort of efficiency constraint  Usually require that several distinct tasks be performed  Key issues include robustly determining which robot should perform which task so as to maximize the efficiency of the team and ensuring the proper coordination among team members

9. Background Two bodies of previous research are particular applicable to “intentional cooperation” Developing control algorithms and implementing them either on physical robots or on simulations of physical robots  Noreils: sense-model-plan-act control architecture  Caloud et al.: another sense-model-plan-act architecture  Asama et al.: decentralized robot system called ACTRESS  Wang: the use of several distributed mutual exclusion algorithms  Cohen et al.: hierarchical subdivision of authority to address the problem of cooperative fire-fighting

10. Background Distributed Artificial intelligence (DAI)  Produced a great deal of work addressing “intentional” cooperation among generic agents  Typically software systems running as interacting processes to solve a common problem rather than embodied, sensor-based robots  Use a distributed, negotiation-based mechanism to determine the allocation of tasks to agents

11. Motivations for ALLIANCE Earlier DAI approaches typically either make no serious effort at achieving fault tolerant, adaptive control or assume the presence of unspecified “black boxes” that continually monitor the environment and provide recovery strategies Control architecture must explicitly address the dynamic nature of the cooperative team and its environment to be truly useful in real-world applications The earlier approaches break the problem into a traditional AI sense-model-plan-act decomposition rather than the functional decomposition used in behavior-based approaches A behavior-based approach to cooperation should be used to increase the robustness and adaptivity

12. Assumptions of ALLIANCE The robots on the team can detect the effect of their own actions, with some probability greater than 0 Robot i can detect the actions of other team members for which i has redundant capabilities, with some probability greater than 0 Robots on the team do not lie and are not intentionally adversarial The communications medium is not guaranteed to be available The robots do not process perfect sensors and effectors Any of the robot subsystem can fail, with some probability greater than 0 If a robot fails, it cannot necessarily communicate its failure to its teammates A centralized store of complete world knowledge is not available

13. Overview of ALLIANCE Design goals: create robot teams that are able to cope with failures and uncertainty in action selection and action execution, and with changes in a dynamic environment A fully distributed, behavior-based software architecture which gives all robots the capability to determine their own actions based upon their current situation No centralized control is utilized Defines a mechanism that allows teams of robots to individually select appropriate actions

14. Overview of ALLIANCE Low-level behaviors, or competences, corresponds to primitive survival behaviors such as obstacle avoidance, while higher-level behaviors correspond to higher goals such as map building and exploring ALLIANCE delineates several behavior sets that are either active as a group or are hibernating Action selection is controlled through the use of motivational behaviors, each of which controls the activation of one behavior set Only one behavior set is active at any point, but other lower-level competences such as collision avoidance may be continually active

15. Overview of ALLIANCE

16. Motivational Behaviors Motivation provides the robots the ability to respond to unexpected events and robot failures Motivational behavior: the primary mechanism for achieving adaptive action selection in the architecture Each motivational behavior receives input from a number of sources The input is combined to generate the output of a motivational behavior The output defines the activation level of its corresponding behavior set When the activation level exceeds a given threshold, the corresponding behavior set becomes active

17. Motivational Behaviors Two types of internal motivations are modeled in ALLIANCE: robot impatience and robot acquiescence The impatience motivation enables a robot to handle situations when other robots fail in performing a given task The acquiescence motivation enables a robot to handle situations in which it, itself, fails to properly perform its task ALLIANCE utilizes a simple form of broadcast communication to allow robots to inform other team members of their current activities The design of the motivational behaviors in ALLIANCE also allows robots to adapt to unexpected environmental changes which alter sensory feedback

18. Motivational Behaviors The parameters controlling motivational rates of robots under the ALLIANCE architecture can be adapted over time based on learning L-ALLIANCE, an extension to ALLIANCE, provides the mechanisms for accomplishing parameter adaptation ALLIANCE architecture is developed to explicitly address the issue of fault tolerance amidst possible robot and communication failures While some efficiency may be lost as a consequence of not negotiating the task subdivision in advance, robustness is gained if robot failures or other dynamic events occur at any time during the mission

19. Formal Model of ALLIANCE Problem definition Let the set R={r 1, r 2, …,r n } represent the set of n heterogeneous robots composing the cooperative team, and let the set T={task 1, task 2, …, task m } represent m independent subtasks which compose the mission  High-level task-achieving function: corresponds to the functions possessed by individual robots  In the architecture, each behavior set supplies its robot with a high- level task-achieving function  The high-level task-achieving functions, or behavior sets, possessed by robot r i is referred to the set A i ={a i1, a i2,…}.  The set of n functions {h 1 (a 1k ), h 2 (a 2k ), …,h n (a nk )} is defined, where h i (a ik ) returns the task in T that robot r i is working on when it activates behavior set a ik.

20. Formal Model of ALLIANCE Threshold of Activation The threshold of activation of a behavior set is given by one parameter θ. This parameter determines the level of motivation beyond which a given behavior set will become active Sensory Feedback Provides the motivational behavior with the necessary information to determine whether its corresponding behavior set needs to be activated

21. Formal Model of ALLIANCE Inter-Robot Communication ρ i : the rate at which robot r i broadcasts its current activity τ i : the period of time robot r i allows to pass without receiving a communication message from a specific teammate before deciding that that teammate has ceased to function Suppression from Active Behavior Sets When a motivational behavior activates its behavior set, it simultaneously begins inhibiting other motivational behaviors

22. Formal Model of ALLIANCE Robot Impatience  Ф ij (k,t): gives the time during which robot r i is willing to allow robot r k ’s communication message to affect the motivation of behavior set a ij  δ_slow ij (k,t): the rate of impatience of robot r i concerning behavior set a ij while robot r k is performing the task corresponding to behavior set a ij  δ_fast ij (k,t): the rate of impatience of robot r i concerning behavior set a ij in the absence of other robots performing the task h i (a ij )

23. Formal Model of ALLIANCE Robot Acquiescence  ψ ij (t): the time that robot r i wants to maintain behavior set a ij activation before yielding to another robot  λ ij (t): the time robot r i wants to maintain behavior set a ij activation before giving up to possibly try another behavior set Motivation Calculation

24. Parameter Settings The parameter settings in ALLIANCE strongly influence the global performance of the system The desirable characteristics of fault tolerance and adaptivity that are present in ALLIANCE should not be sacrificed while enabling increases in robot team efficiency L-ALLIANCE: provides mechanisms that allow the robots to dynamically update their parameter settings based upon knowledge learned from previous experiences Assumptions  A robot’s average performance in executing a specific task over a few recent trials is a reasonable indicator of that robot’s expected performance in the future  If robot r i is monitoring environmental conditions C to assess the performance of another robot r k, and the conditions C change, then the changes are attributable to robot r k

25. L-ALLIANCE Incorporates the use of performance monitors for each motivational behavior Robot r i programmed with the b behavior sets A={a i1,a i2,…,a ib }, also has b monitors MON i ={mon i1,mon i2,…mon ib } Monitor mon ij observes the performance of any robot performing task h i (a ij ) Monitor mon ij uses a mechanism to update the control parameters of behavior set a ij based on the learned knowledge

26. Action Selection Algorithm

27. Parameter Settings

28. Results The ALLIANCE architecture has been successfully implemented in a variety of proof-of-concept applications on both physical and simulated mobile robots Over 60 logged physical robot runs of the hazardous waste cleanup mission and over 30 physical robot runs of the box pushing demonstration were completed to elucidate the importance issues in heterogeneous robot cooperation

29. Hazardous waste cleanup The Robots  Three R-2 robots purchased commercially from IS Robotics  Mechanical drift and failure can cause them to have quite different actual abilities  A radio communication system allows robot team members to communicate with other The Mission  Two artificially “hazardous” waste spills in an enclosed room to be cleaned up  The distinct tasks: locating the two waste spills [find-locations]; moving the two spills to a goal location [move-spill (left) and move-spill (right)]; periodically reporting the team progress to humans monitoring the system [report- progress]

30. Hazardous waste cleanup Behavior sets:  find-locations- methodical  find-locations-wander  move-spill (loc)  report-process  avoid-obstacles

31. Experiments Three robots are referred as GREEN, BLUE, and GOLD

32. Experiments: no robot failure RP: report-progress MS(L): move-spill (left) MS(R): move-spill (right) FLW: find-locations-wander FLM: find-location-methodical

33. Experiments: interfere with GREEN

35. Experiments

37. Experiments: removal of BLUE

39. Discussion The cooperative team under ALLIANCE control is robust The cooperative team is able to respond autonomously to many types of unexpected events either in the environment or in the robot team without the need for external intervention The cooperative team needs not have a priori knowledge of the abilities of the other team members to effectively accomplish the task The primary weakness of ALLIANCE is its restriction to independent subtasks

40. Conclusion ALLIANCE is a fully distributed, behavior based architecture which facilitates fault tolerant mobile cooperation ALLIANCE allows the robots to handle the environmental changes fluidly and flexibly ALLIANCE also allows robot team members to respond to their own failures or to failures of teammates, leading to adaptive action selection to ensure mission completion ALLIANCE further enhances team robustness by making it easy for robot team members to deal with the presence of overlapping capabilities on the team

41. Final word Thank you and Have a great new week!