Mental Models for Human-Robot Interaction Christian Lebiere (cl@cmu.edu)1 Florian Jentsch and Scott Ososky2 1Psychology Department, Carnegie Mellon University 2Institute for Simulation and Training, University of Central Florida Cognitive Models of Mental Models • Mental models provide a representation of situation, various entities, capabilities, & past decisions/actions • Current models are non-computational descriptions • Cognitive models can provide computational link to overall robotic intelligence architecture for dual uses: – Provide a quantitative, predictive understanding of human team shared mental models – Support improved design of human-robot interaction tools and protocols – Provide a cognitively-based computational basis for implementation of mental models in robots Representation Components • Mental model representation – Ontology of concepts and decisions • Lexical (WordNet), Structural (FrameNet), Statistical (LSA) – Symbolic frameworks • Decision trees, semantic networks – Statistical frameworks • Bayesian networks, semantic similarities • Knowledge of task situation – Situation awareness – mapping to levels of SA – Environment limitations – who sees/knows what (perspective) – Architectural limitations – who remembers what (WM, decay) Reasoning and inference • Inferring mental models – Instance-based learning (Gonzalez & Lebiere) • E.g., Learning to control systems by observation or imitation • Inferring current knowledge – Perspective-taking in spatial domain (Trafton) • E.g., hide and seek, collaborative work • Predicting decisions – – – – Theory of mind recursion (Trafton, Bringsjord) Imagery-based simulation (Wintermutte) Shared plan execution in MOUT (Best & Lebiere) Sequence learning in game environments (West & Lebiere) Cognitive Architectures – Production systems – Utility – rewards and costs • Memories – Working memory: buffers – Long-term: semantic/episodic – Activation mechanisms • Learning – Symbolic and statistical Intentional Module (aPFC) Declarative Module (Temporal/Hippocampus) Goal Buffer (DLPFC) Productions (Basal Ganglia) • Computational representation of invariant cognitive mechanisms • Behavior selection Retrieval Buffer (VLPFC) Matching (Striatum) Selection (Pallidum) Execution (Thalamus) Visual Buffer (Parietal) Visual Module (Occipital/etc) Manual Buffer (Motor) Manual Module (Motor/Cerebellum) • Human factor limitations – Perceptual-motor parameters • Individual differences – Strategies and knowledge – Capacity parameters Environment Pursuit Task • Follow that Guy: human soldier and robot teammate – Shared mental model of pursuit situation scenario • Set of data encoding various scenarios • Items organized according to SMMs held by expert teams (Equipment, Task, Team Interaction, Team) • Decision tree built using information from police “foot pursuit” procedures • For each decision, the most critical item is listed – However, other factors may be considered in weighing decision • Loop to end or continue the pursuit given fluid situation Data Model Structure Model Information [Code] Data Type Example Scenarios Unit Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Scenario 7 Scenario 8 Equipment Model (EQ) Communications EQ-C1 GPS enabled for location transmission? binary 1 1 1 1 1 0 1 1 EQ-C2 Radio frequency integer hz 250 263 250 263 250 250 250 263 EQ-C3 Radio coverage integer Miles 5 5 3 5 5 2 5 0 EQ-S1 Optical Range meters 60 60 60 60 60 60 60 60 EQ-S2 Heat Sensing binary no no no no no mp no no EQ-S3 Infrared (3D mapping) binary yes yes yes yes yes yes yes yes Sensor Equipment Weapon Systems EQ-W1 Weapons platform binary no no no no no no no no no EQ-W2 Robot weapon type data set range, caliber, clip size n/a n/a n/a n/a n/a n/a n/a n/a EQ-W3 Soldier weapon 1 (M16A2) data set range, caliber, clip size 550m, 5.56mm, 30 550m, 5.56mm, 30 550m, 5.56mm, 30 550m, 5.56mm, 30 550m, 5.56mm, 30 550m, 5.56mm, 30 550m, 5.56mm, 30 550m, 5.56mm, 30 EQ-W4 Soldier weapon 2 (M9) data set range, caliber, clip size 50m, 9mm, 15 50m, 9mm, 15 50m, 9mm, 15 50m, 9mm, 15 50m, 9mm, 15 50m, 9mm, 15 50m, 9mm, 15 50m, 9mm, 15 59 suspicion of criminal mischief 187 suspected attempt to murder, Suspect SK-S1 Reason for pursuit integer SK-S2 Is (are)the suspect(s) armed? binary SK-S3 Whether ID is known binary SK-S4 What is the severity of the offense? scale of 1-10 rating SK-S5 Age integer age Description of suspect(s) SK-S6 What direction is suspect(s) heading in? SK-S7 police code? 417 Person with gun, 10-71 shots 59 suspicion of criminal mischief Small rocket explosion strikes 171 Suspicion of trespassing, 467 147 suspected assult with a deadly 246, shooting at inhabited dwelling. fired 171 Suspicion of trespassing 59 suspicion of criminal mischief 211 , Robbery nearby building possession of deadly w eapon weapon 1 1 1 0 0 1 1 1 0 0 0 0 1 0 0 0 8 8 5 4 6 8 10 10 27-33 16-20 21-26 30-35 20-25 null 35-40 20-25 Male, Cau casian, visible bodily Male, Cau casian, visible bodily data set facial data, clothing color, m/f, etc. markings, tattoo. Male, black, 5'10, 150 lbs. Male race unknown, 5'9, 165 lbs Male, Arabic, 6', 170 Male, Arabic, 5'4, 150 null Male race unknown, 5'9, 165 lbs markings, tattoo. integer degree / h eading 0,45 180, 225 0,45 180, 225 0,45 null 0,45 180, 225 Task Model (SK) Is (are) suspect(s) an immediate threat to public SK-S8 safety? binary 1 1 1 1 0 1 1 1 SK-S9 Presence of additional suspect(s)? binary 0 0 0 0 0 1 0 0 SK-S1 0 Fitness level of suspect scale of 1-10 7 9 6 7 5 0 7 6 SK-S1 1 Is the suspect(s) injured? binary 0 0 0 0 0 0 0 0 SK-S1 2 Severity of injury scale of 1-10 rating n/a n/a n/a n/a n/a n/a n/a n/a scale of 1-10 severity 1 2 4 2 1 1 1 8 rating Environment SK-E1 Weather conditions SK-E2 Amount of vehicular or pedestrian traffic (density) scale of 1-10 density rating 1 8 4 1 9 3 1 4 SK-E3 Condition of traveling surfaces scale of 1-10 difficulty rating 7 rough, rugged 2 street scape, paved 5 Street scape, fair 10 Street scape, poor 1 Street scape, good 2 street scape, paved 1 Street scape, good 5 Street scape, fair SK-E4 Nature of area data set Marketplace, industrial, etc. Rural, Wooded Residential Residential Commercial Residential Industrial Residential Residential Condition of structures (e.g., abandoned, SK-E5 condemned) integer pre-coded Natural inhabited inhabited Abandoned inhabited abandoned inhabited inhabited SK-E6 Area visibility integer miles 3 2 1 2 2 4 2 1 SK-E7 (Human) familiarity with the area scale of 1-10 rating, if applicable 2 6 9 5 7 5 4 6 SK-E8 Time of day time time 13:00 18:52 15:45 7:10 2:32 14:40 18:05:20 13:24:58 SK-E9 Hostile Environment? binary 0 0 0 0 0 1 0 0 Availability alternative scenarios SK-A1 Apprehension at another time binary 0 0 0 0 1 0 0 0 SK-A2 Perimeter to contain suspect(s) binary 1 1 1 1 0 1 1 1 SK-A3 Aerial support binary 0 1 0 0 0 1 0 1 SK-A4 Canine search binary 1 1 0 0 0 0 1 0 SK-A5 Saturation of the area with patrol personnel binary 0 1 0 1 0 1 0 0 Team Interaction Model (IA) Interaction patterns IA-A1 Are backup units available? binary 1 1 1 1 0 1 1 1 IA-A2 Has communications been notified? binary 1 1 1 1 1 0 1 1 IA-A3 Can a perimeter be set up? binary 0 1 1 1 1 0 1 1 IA-A4 Line of sight with teammate? binary 1 1 1 1 1 0 1 1 1 1 1 1 0 0 1 0 Roles and Responsibilities IA-R1 Has clearance been given by supervisor? Binary IA-R2 Who is responsible for calling off the pursuit? data set rank, autonomy level Supervisor Supervisor Supervisor Supervisor Supervisor Warfighter Supervisor Supervisor data set rank, autonomy level Warfighter Warfighter Warfighter Warfighter Warfighter Warfighter Warfighter Warfighter data set coded protocol instruction set Permissive Conservative Conservative Permissive Conservative Hostile Conservative Conservative Who is responsible for callin g communications/supervisor? IA-R3 Mission protocols (permissive, hostile, aggressive, conservative, etc.) IA-R4 Team Model (TM) Warfighter (P1) TM-W1 Do I have visual contact with the suspect? binary 1 1 1 1 1 0 1 1 TM-W2 Physical condition of the warfighter scale of 1-10 fitness, injury rating 8 8 8 8 8 8 8 8 TM-W3 Knowledge and skills of warfighter data set rank / class based data Private 1st class, M.P. Trained Private 1st class, M.P. Trained Private 1st class, M.P. Trained Private 1st class, M.P. Trained Private 1st class, M.P. Trained Private 1st class, M.P. Trained Private 1st class, M.P. Trained Private 1st class, M.P. Trained TM-W4 Years of experience integer years 3 3 3 3 3 2 3 3 TM-W5 Possession of a firearm? binary 1 1 1 1 1 1 1 1 TM-W6 Personality characteristics data set profile data of teammates Robot (UGV 1) TM-R1 Terrain traversal rating integer for dirt, gravel, bricks, etc. (1-10) 5 5 6 5 6 5 6 5 TM-R2 Average speed integer mph 5 5 5 5 5 5 5 5 TM-R3 Top speed integer mph 8 8 8 8 8 8 8 8 TM-R4 Robot platform armed? binary 1 1 1 1 1 0 1 1 TM-R5 Platform armor rating integer code 6 6 6 6 6 8 6 6 TM-R6 Maximum useful sensor range integer ft 1000 1000 1000 1000 1000 500 1000 1000 TM-R7 Energy level remaining time hh:mm:ss 12:45:37 8:32:05 6:01:21 2:48:13 10:06:47 3:00:00 10:21:47 6:04:23 TM-R8 Damage assessment scale of 1-10 rating 1 1 1 1 1 5 1 1 4 1 3 2 2 3 1 4 1 3 2 4 3 1 2 4 2 4 3 1 3 2 1 4 2 3 1 4 3 1 2 4 Team responds to gunfire in the area Team responds to gunfire in a You and your partner are You and your partner are sent to You and your teammate are While patrolling an industrial area You and your teammate are You and your teammate are during the afternoon. Civilian safety in residential area in the evening. performing a routine afternoon perform an inspection of what patrolling jus outside of the Action selection 1 2 3 4 Robot pursues Suspect, Soldier holds position Soldier pursues Suspect, Robot holds position Soldier and Robot pursue Suspect together Robot and Human do not pursue, file report Strategy Rank Order (1 is the best solution) Scenario Synopsis (for descriptive purposes) plagued by intermittent fighting, a performing a street patrol a few performing a routine afernoon this area is a concern. A suspect is Civilian safety is a concern. A suspect patrol of a residential area in appear to be abandoned buildings military base situated within city nearby blast strikes an abandoned minutes before a curfew is about patrol of a residential area when spotted that fits the reported is spotted matching the description. which some of the residents have to sweep the area for IEDs. It is limits. You see a robbery take builting. There are no casualties, you hear shots fired. Upon description. However, identity is not Suspect flees on foot. Suspect takes been suspected of aiding a around 7am, a few hours before a place, but the suspect does not however witnesses report seeing a appears to be an abondoned approaching the scence you see a known at this time. Upon detection, off running down a crowded street. terrorist organization members. civilian construction team is appear to be armed. The suspect rocket or mortar fired from a crowd over people standing over suspect flees on foot. Suspect runs into This is a busy evening and there are Upon rounding a corner you see a scheduled to start working on the appears to be a man that you nearby building window. The robot cautiously. As you do, you see a an injured man and another man a wooded area in which the terrain is many civilians present. The sun has man start running away from you. buildings. When you approach the stopped last week and is well has sustained moderate damage, man with what looks like a radio running away. A severe storm is difficult for your robotic partner to set and it is starting to get dark You notice that he is carrying a known in the area. It's late including optical sensor overload controller and a handgun. He moving in and it is hindering the maneuver through. However, the outside. However, the weather gun. You have never seen this man running down the street an away afternoon and the streets are full and gps failure. Mobility and other notices that you have spotted him ability of your radio equiptment to direction the suspect is heading is away conditions are clear and warm. before and his identity is from you. You call out to him to of civilians. Weather and traveling sensors appear unharmed. It is the and starts to run in the opposite from the town center. Weather Additional units approximately 2 unknown. You and your partner stop, but he continues running. conditions are good. However, all middle of the day, clear skies, and direction . It 's getting late in the teammate. Otherwise, road conditions are clear and warm. The miles away with an ETA of 5 minutes. are familiar with this area and The area is not heavily populated of the potential backup units are backup patrols are naerby in the evening and most of the civilians surfaces are in fair condition and unavailable at this time. area. in the area are in observance of pedestrian and vehicluar traffic is the curfew. The wheather is clear moderate in this area. Backup units are available to help set up a building, you see a man start to start. You come across what vehicle. You approach the vehicle communicate with your robitc closet available backup unit is a k-9 unit backup units are close by that can with civilians. However, the that is approximately 3 miles away, ETA help you contain the suspect. 7 minutes. Weather conditions are cloudy but poor condition. Weather conditions and the road surfaces are in good do not present a threat at this condition. In addition, there is a k- parimeter in the area. traveling surfaces and roads are in are clear and there is another team Scenario Data and Decision Tree Part 1: Who should pursue? Start H-R Communication reliable (5x5)? YES Is the terrain negotiable for robot? EQ-C3 No Are sensors reliable in the search area? YES SK-E3 No YES Soldier only pursuit No Immediate threat / critical situation? IA-A1 Are suspects armed? EQ-S3 No Is backup support available? YES SK-S2 SK-S8 No No Current last known location? No No SK-S7 SK-S8 YES YES Hold position, report incident Is the threat immediate (civilians, etc.) YES Robot only pursuit Team pursuit Continue to Part 2: pursuit loop YES Deciding whether to pursue Was this, or is there potential for a violent crime? SK-S1 No Is the suspect armed? SK-S2 Yes Yes No No No Yes Can you apprehend them at a later time? SK-A1 Yes Yes Are communications functioning properly? EQ-C3 Are backup units available to assist you? IA-A1 Discontinue and Report Do you know the identity of the suspect? SK-S3 No No Can a perimeter be set up to contain the SK-A2 suspect? Yes No Yes Do you have supervisor clearance? IA-R1 No Begin or Continue pursuit Do you have line of sight with suspect? TM-W1 No Yes What are the weather conditions? SK-E1 Poor Good/ Fair Continue Pursuit Yes What is the pedestrian traffic like? SK-E2 Heavy Light/ Moderate What are the traveling surface conditions? SK-E3 Poor Good/ Fair General Cognitive Model • Develop general model that takes mental models in the form of decision trees and learns to retrieve and execute them • Each decision is represented as sequence of chained steps • Each piece of data is represented as separate chunk • Model (7 p* production rules) depends on declarative memory to retrieve rule steps, data items and decision instances – No hardcoded decision logic • Each decision depends on matching against past instances combining activation recency, frequency and partial matching • Stochasticity of activation results in probabilistic decisions • Run model in Monte Carlo mode for decision distribution • Cross-validation: train on some scenarios, test on others Individual Decision Inference 1 Probability "Yes" Decision 0.9 0.8 0.7 0.6 0.5 Factor1 = yes 0.4 Factor2 = no 0.3 0.2 0.1 0 zero one two three Factor 2 Value four five Overall Decision Agreement 1 Decision Probability 0.1 0.01 Scenarios Average 0.001 0.0001 SUTR TR SU R T SU S TRU UR T S T S RU Ranked Decisions TR SU S TRU 1 2 3 4 Generalized Condition • 35 scenarios • 3 experts • Intermediate decisions • Relative rankings • Desirability ratings • Comments Results 0.6 0.5 • Match to first-last ranks, poor middle • Slightly different ratings pattern • Comparable crossexpert correlations 0.4 Model Rank 1 Model Rank 2 0.3 Model Rank 3 0.2 Model Rank 4 0.1 0 Human Rank 1 Human Rank 2 Human Rank 3 Human Rank4 0.8 0.7 0.6 0.5 Human Desirability Ra ng 0.4 1-vs-2 Human Ratings Model Probs Human Rankings Model Rankings 1-vs-3 0.37 0.19 0.43 0.33 2-vs-3 0.31 0.46 0.21 0.39 0.2 0.09 0.24 0.05 Model Probability 0.3 0.2 0.1 0 0 1 2 3 4 5 Learning • Proceduralize individual steps from declarative instructions to production rules to replicate learning curve from novice to proficiency and expertise • Apply feature selection using utility learning to encode and use only a subset of data items for each decision • Learn shortcuts that combine multiple individual binary decisions into single, multi-outcome decision • Generate rankings/ratings from probability judgments generated from activation of memory retrievals • Abstract decision instances into discrete types Future Work • Validate model against human participants data along entire learning curve and broad range of situations • Explore Bayesian network formalism as alternative to enhance generalization in multi-step decisions • Integrate cognitive model in multi-agent simulations to validate computational mental model in dynamic decision-making setting • Integrate computational cognitive model on robotic platform to assess ability to improve human-robot interaction through shared models