Emotion-Based Control of Cooperating Heterogeneous Mobile Robots

1. University of Missouri – Columbia Emotion-Based Control of Cooperating Heterogeneous Mobile Robots Robins R. Murphy, Christine L. Lisetti, Russell Tardiff, Liam Irish and Aaron Gage Presented by Ashwin Mohan Course Instructor Dr. Majorie Skubic April 27, 2005

2. What does this paper talk about? • First investigation a formal cognitive model of emotions – choice of behavior is self-regulated for each agent • Investigates the capacity of agents to distinguish and adapt based on emotions • Investigates ability of robot to represent and learn knowledge using emotions to improve efficiency • Investigates how emotions enable adaptation to harmful conditions • Uses communication, awareness of and reaction to emotional stimuli to achieve interdependency

3. Why interested in INTERDEPENDENCY? Multi-agent control for interdependent tasks : “Execution of a tightly coupled task with a cyclic dependency and one that cannot be performed by another robot” • Interdependency leads to the possibility of failure – hardware, planning, environment, etc • How can one enhance performance improvements in a dynamic and distributed system • What kind of control mechanisms can be implemented on a heterogeneous team

4. Competition, Interference and Deadlock • Robots cooperating asynchronously on a sequential task can enter deadlock, where one robot does not fulfill its obligations in a timely manner • Interference includes conflicts like goal clobbering, deadlocks and oscillations • Resource competition can be over space, information, and objects • Stagnation occurs when a team of robots work on a task but cease to make progress

5.  Motivation for using Emotions  • One method of controlling multi-robot systems • Help to dynamically adapt to limitations, manage social behavior, and to communicate with others • At the implementation level, emotions monitor the accomplishment of the goals and corresponds to a formal cognitive model • Emotional intelligence lead to robots capable of representing and learning affective knowledge • Each emotion calls a distinctive suite of actions appropriate for that emotional state

6. Related work and non-cognitive solutions • Coherent framework for implementation of emotions in heterogeneous robots is missing • Based on structure of behaviors, NO awareness of interaction • All tasks were available to all robots • Work focused on functions of emotions in social exchanges than efficiency • Very few AI models have computer programs • No explicit representation and awareness of and reaction to emotional stimulus • No work done on interdependence in hybrid architectures

7. Framework for this paper • Leventhal and Scherer:the hierarchical multilevel process theory of emotions • The sensory-motor levelis activated by external stimuli and internal changes. Reactions are mostly of short duration and reflex-like • The schematic levelintegrates sensory-motor processes with prototypes of emotional situations having concrete representations • The conceptual levelis deliberative, involves reasoning over the past and projecting into the future to avoid emotional disturbances

8. Approach using BSG and ESG • Scripts used for assemblages of robotic behavioral schemas to represent stereotypical set of behaviors • BSG and ESG accept measures of task progress as inputs • Task progress metrics come from three sources: monitors, individual behaviors, and inter-agent communication • Monitors are perceptual schemas; Individual behaviors often act as releasers for other behaviors, communication from an external agent, either a command (e.g.,“hurry”) or data (e.g., “I’m at location ”) Fig A : Layout of a causal chain

9. Approach (2) • The use of emotions and ESG breaks the potential master/slave coupling Example: • Robot A receives a message “Hurry” • Causes shift, say from “Confident” to “Concerned” • ESG triggers adaptation to make better progress • Robot may reduce the sensitivity to obstacles, change the acceleration of its motions to produce the change

10. Multi Agent Implementation • Waiter Script implements only schematic link • Refiller Script implements only the sensory-motor level link between ESG and behaviors • BSG and ESG represented as a single finite state machine Fig B : Basic organization of the reactive layer of Sensor fusion effects (SFX)

11. Waiter robot whose task it is to serve items to an audience Refiller cannot perform the Waiter’s task and vice versa Two robots are distributed and decentralized Each robot uses WaiterScript or RefillerScript Refiller can substantially decrease the time on task, i.e., serving The tray of refills is the resource Emotions are used to adapt or change the team behavior Robots use wireless to communicate either a command or location data Problem setup

12. Fully autonomous Nomad 200 robot bases Both robots use the (SFX) hybrid deliberative/ reactive architecture Both robots run under RedHat Linux version 3.0.3 and are coded in a combination languages Emotional state was governed by the changing relationship of the rate of treat consumption, time till empty (TTE) to the time to be refilled (TTR) The Waiter (a.k.a. Butler) has sonar rings, laser ranger, a thermal probe, and dual Hitachi color video camera and controlled by on-board processors (233-MHz and 133-MHz MMX) The Refiller (a.k.a. Leguin) has one sonar ring, and dual Hitachi color video cameras on a pan-tilt head and has 233 and 66-MHz Pentium MMX’s Knowledge Query and Manipulation Language (KQML) agent communication language Problem setup (2)

13. Action tendency and Task progressmeasures • Robots programmed with the happy, confident, concerned and frustrated emotional states which correspond to the action • Two modifiers were used, caution, C, and patience, P, acting as thresholds • The output of the emotions was at the schematic level, leading to changes in the set of active behaviors Fig C : Action Tendency Fig D: Task Progress Measures

14. Waiter and WaiterScript • One external input data about location of Refiller • Six external outputs which are communicated to the Refiller • Wait & refill, produced by serve • Hurry, intercept, go home generated when an emotional event (state change) occurs • script is responsible for computing (TTR) & (TTE), task progress measures and instantiating or modifying the set of active behaviors Fig E : WaiterScript

15. Refiller and the RefillerScript • One source of inputs from the Waiter as commands (wait, refill, hurry, go home, intercept) or position data • Wait is when she loiters around the serving station • Refill is a move-to-goal behavior where Waiter is the goal • Hurry command increases her navigational speed to the maximum safe speed Fig F : RefillerScript

16. Combined Behavior Diagrams Fig G : Representative data run showing emotions and changes in internal and team behavior

17. Combined Behavior Diagrams (2) • WaiterScript begins with the serve behavior Happy and sends a command wait to the Refiller to synchronize • Serve uses the sub-behavior face-find to track human faces plays sound encouraging to remove treats • While serving, at t150 the Waiter robot may communicate a “refill” request if (TTE) is now less than (TTR), & Confident that she will receive a refill in time • Refill changes behavior from wait to refill • Ideally, the Refiller reaches the Waiter triggering the exchange behavior • At about 220s, Waiter goes into Concerned triggering hurry to Refiller

18. Combined Behavior Diagrams (3) • Insufficient progress of Refiller will cause emotional state of Frustrated • ESG now dictates, if Butler should abandon serve and move to intercept • During intercept, emotional state moves from Frustrated to concerned to happy • Under exchange, the Waiter does nothing until the tray-watch monitor sees the operator flash the empty tray in front of the cameras • When the tray was seen, the Waiter communicates a “go home” command to the Refiller. • When exchange terminates, the WaiterScript re-instantiates serve

19. Discussion of Results • Demonstrated at AAAI Mobile Robot Competition,’00, in Austin, TX, and at Museum of Science and Industry (MOSI), 2000 • Trace data collected at MOSI clearly showed that emotions led to dynamic adaptations, and changes in the robots’ behaviors • Robots regulated their subgoals and motivations according to their own current internal emotional states as well as external signals • They socially adapted their actions to the other agents, both human and artificial, depending on the current situation • Emotional model provides number of features beyond a simple state machine • Coding of emotions is simple and can be added to these systems without any re-conceptualization of components

20. Discussion of Results • Emotional implementation is local to each robot and based on task progress • Robots need not interpret or understand each other’s emotions • Sensory-motor stimuli (e.g., seeing a tray full/empty) gave rise to simple reflex-like reactions involving the motor system only • Emotions suited for control of distributed, behavior-based systems where centralized, deliberative methods are often too computationally expensive • The implementation reported in this paper is appropriate for behavior-based and hybrid deliberative/reactive robots • Partial translation of the multilevel process theory of emotions • Results offers support that this multilevel theory is a useful model of emotions

21. Thank you!