Hybrid Human-Agent Teams for Simulation Based Training 1 / 53 Part 1 – The VR Environment 2 / 53 Background • The MRE (Mission Rehearsal Exercise): Aims “…to create a virtual reality training environment…” in which various scenarios that may be encountered by military units on peace-keeping missions can be played out. 3 / 54 4 / 54 Introduction Methodology in the design of the system: • Selecting state-of-the-art systems for different needs of the simulation. • Improving/modifying these systems where needed. • Developing specialized software for the simulation (the actor agents, and their AI). • Combining these different systems as seamlessly as possible. 5 / 54 Tweaking • Step-by-step improvements of the different components of the system strive to create a more believable and realistic experience for the trainee. • Every little step forward can take us a long way towards better virtual reality. 6 / 54 System Components • Video: – Panoramic (150°), semi-circular, 2 2/3 meter high screen encompasses the user. – Environment and special effects rendered using commercial Vega™ program. – Characters are rendered using the commercial PeopleShop™ program. – Rendering is done in real-time. 7 / 54 System Components (2) • Audio: – Complex multi-channel audio system, with speakers in front, sides, rear and overhead. – Provides all background sound, with strong spatial capabilities. – High realism. – Large efforts were made to synchronize sound and events occurring on the screen. 8 / 54 Human Interface • Input (speaking to characters) – Speech recognition – Identifies keywords – No deeper understanding – Plans for the future include merging natural language technologies. 9 / 54 Human Interface (2) • Output (Agents to humans) – Pre-recorded statements – Synthesized speech (in real-time) – System can produce nuances in synthesized speech, adds to believability. – Lips are synchronized with generated speech! 10 / 54 Eliminating know-it-alls • What would happen if the mother agent could know everything you whispered to your sergeant? • What if your sergeant criticized your actions based on facts you he had no way of knowing? 11 / 54 Limiting Perception • Filters were placed on the agents perceptions, so that they only perceived the parts of the simulation that a normal person would. – Don’t have eyes behind their back – Can’t hear things far away, or with too much background noise (helicopter flying over). The sergeant will cup his ear when he can’t hear you over the noise!! 12 / 54 The Script • We can’t allow the trainee to do whatever he/she wants, both because of limits in technology, and because of the training goals of the system. • On the other hand, if every attempt to deviate from the path is punished, the user will soon get frustrated, and won’t actually learn to find the right solution. 13 / 54 StoryNet • The StoryNet model allows several plot nodes, with transitions between them. • Within each node, there is a relatively area of ‘freeplay’. • Transitions between nodes are caused by key actions or events, in a way transparent to the user. 14 / 54 15 / 54 The Director • In the future, there are plans to implement a Director agent, who will be in charge of guiding events to the correct path, if the student deviates too much. 16 / 54 Last but not least – the Cast • Three types of Agents/actors: 1. Scripted : have predetermined behavior (movement/actions) which is triggered by the controller of the simulation, or by the actions of other agents. 17 / 54 The Cast (2) 2. AI : these include the main characters in the simulation, with which the user has direct contact. These agents have a set of general, domain independent capabilities, operating over a declarative representation of domain tasks. The task are represented to the agents as a set of steps (primitive actions, or other tasks), a set of ordering constraints, and a set of causal links that describe what is accomplished by each step, and why it is needed (to achieve a precondition for another step). 18 / 54 • This ‘knowledge’ about the world and tasks in it, allow the AI agents to: – generate a plan to fill their tasks, – change their plans when the world state changes, – and maintain a dialog with humans and teammates about the task. 19 / 54 The Cast (3) 3. AI agents with emotions: – Are built on top of the regular AI model. – Express emotion with gestures, and nuances of speech (!!). – Use their plan model to decide which emotions to express. – The emotional state influences the agent’s beliefs, desires and intentions. 20 / 54 Expressing Emotions • The emotional agents use their plan model to decide which emotions to express. 21 / 54 Expressing Emotions (cont.) • If the mother believes that the commander plans to leave, and that that will make her task of helping her child impossible, she may express anger. • If she believes the commander plans to stay, she may express hope. 22 / 54 Emotions (2) The emotional state influences the agent’s beliefs, desires and intentions: • The sergeant is under stress because he’s responsible for the boy’s injury. • He may seek to relieve the stress by – shifting the blame to others, or by – • forming an intention to help the boy get medical help. This choice will depend on other, adjustable, character traits/states-of-mind. 23 / 54 Part 2 – Dialogue 24 / 53 Layout • • • • Structure of AI model Domain representation Planning Dialog actions – Speech acts – Grounding – Stances – Negotiation proceedings – Future work 25 / 54 AI Model • The agents use a general reasoning algorithm to ‘understand’ and plan in a declarative representation of team tasks. • The algorithm is a domain-independent non-linear STRIPS planer. • The task representation is domain specific, and encodes the given training scenario. 26 / 54 Task Representation • Each task is represented by a set of steps, which are either – A primitive action – physical or sensory. Or – An abstract action – a sub-task (which in turn is composed of a series of steps, etc.). 27 / 54 Task Representation (2) • In addition, there may be ordering constraints between the steps. • There can be interdependencies between steps, represented by causal links and threat relations. 28 / 54 Task Interdependency • Causal links – specify that one step’s goals are another’s precondition. • Threat relations – step A is a threat to step B, if the completion of A causes the removal of one of the preconditions of B. 29 / 54 Refinements - Responsibility and Authority • Task steps are associated with the team members that are responsible for performing them, and, optionally, with the agent who has authority over that step • The agent responsible for the step should not perform it until authorization is given by the teammate with that authority. 30 / 54 Planning • When the team is given a (top-level) task, each member recursively constructs a complete detailed task model. • Each member builds the model according to his/her knowledge, which may be incomplete or erroneous. 31 / 54 Re-planning • The Agent’s perception module constantly monitors the world, and sends messages to the cognition module about changes in the world state. • Whenever changes occur, the agent rechecks his full plan, and corrects it according to the changes. • For instance, if a goal in the plan has been achieved by a teammate, there is no longer a need to perform the appropriate task yourself. 32 / 54 Re-planning (cont.) • Instead of creating the whole task plan from scratch whenever a change occurs, the agent uses it’s previous model as a guideline, and so cuts down on the amount of alternatives it calculates. 33 / 54 Agent’s Task Model • At any given time, each agent holds a complete plan that models the way in which he believes the whole team can accomplish the task. • This plan allows the agent to answer questions and reason about the task. 34 / 54 Courses Of Action (COA) • In order to negotiate, there is a need to support reasoning over several possible COAs. • Each COA is a high level task, as described above, but is not automatically marked as ‘intended’ by the agent just because it achieves the goal. • Instead, if the agent reasons that the COA can achieve the goal, it is marked as ‘relevant’. 35 / 54 COAs (cont.) • Since different COAs may be mutualy-exclusive, threats and causal links between alternative COAs must be ignored. • On the other hand, the effects of a COA on other tasks are taken into account when calculating the utility of the that COA. • A complex utility function is used to decide between the alternative COAs. 36 / 54 Dialog • The information state is the information representing the complete current context of the dialog. • This model is more complex than the “dialog state” of the standard plan-based approach. • It contains all the necessary data relevant to the current state of the dialog, and data needed to decide on future dialog acts. • The information state is updated via dialog acts, according to dialog rules. 37 / 54 Speech Acts • Core speech acts – influence the topic under discussion – Establish and remove commitments. 38 / 54 Speech Acts • • • • • Assert Info-request Suggest Request Order These acts reflect social commitments of the speakers, but do not directly affect the BDI of the person addressed, since they may be insincere. 39 / 54 Types Of Speech Acts • Speech acts can describe three things: • Action – consists of – Agent : the person performing the action – Event : the action being performed – Patient : the object on which the action is being performed. • State – indicates the state of some aspect of an object. (for instance, whether the door of the house is open or closed). • Question - about one of the above. 40 / 54 Grounding • In order for an act to be considered as accepted into the common ground of the conversation, it needs to be acknowledged by the agent to whom it was addressed. • This is accomplished by grounding acts. • Core speech acts are not seen as having their full effects on the social state until they are grounded. 41 / 54 Grounding Acts • • • • • • • • Request-acknowledge Acknowledge Request-repair Repair Initiate Continue Display Cancel 42 / 54 Stances • The current state of team negotiation is represented by a sequence of stances. • Each stance represents the outward representation of an agent’s view on the issue. 43 / 54 Stances (2) • Each stance is composed of – The agent who holds the stance – The action that the stance is about – The attitude (stance) of the agent towards the action. – The audience to whom the agent is presenting the stance. – The reason for the stance – The time at which the stance was made. 44 / 54 Stances (3) • The actual attitude (stance) can be one of the following (from most positive to most negative): – Committed ()התחייבות – Endorsed ()מומלץ – Mentioned – Not mentioned – Disparaged (שלא-)מומלץ – Rejected ()דחיה 45 / 54 Negotiation • The sequence of negotiation stances shows the progression of the negotiation: progresses from proposals of action towards commitments. • Stances arise from core speech acts (for example: if A requests B to do something, A’s stance is committed to that action). • There are also special negotiation acts: accept, reject, counterproposal, and explanation (for or against), each of which give rise to the appropriate stance. 46 / 54 Initiating Negotiation • An agent’s initiative characteristic determines whether the agent is likely to start a new negotiation. • If another agent starts a negotiation with an order or request, the agent addressed is required to respond in some fashion. 47 / 54 Appropriate Responses • The response to an action proposal is guided by the following factors: – The relevant party : the agent who is in charge of the next step in performing the action (the authorizor, the agent responsible for the action, or the addressee himself) – The dialog state : one of discussed, needsdiscussion or unmentioned. – The plan state: how the action relates to the agents plan (good/bad , intended/not-intended etc.) 48 / 54 Appropriate Responses (cont.) Examples: • The agent will reject the proposal if plan-state is one of { bad, considered-bad, unknown, conflict, goals-satisfied } dialogue-state=needs-discussion. • The agent will accept (reluctantly)! If relevant-party=me plan-state=considered-bad dialogue-state=discussed 49 / 54 Appropriate Response (cont.) • There are possibilities of overlap of conditions, in which case several responses are possible. • The agent has to decide which responses to choose, and in what order. • The decision is based on practical considerations (which is the best, or most immediately relevant), and on social relations between the proposing party and the agent (superiority/rank). 50 / 54 Resolution • Negotiation between teammates proceeds until both agree to accept/reject, or until one drops the contrary stance. • It is also possible to “agree to disagree”, assuming it is possible to proceed with other actions despite the disagreement. 51 / 54 Team Action • In order for an action to be performed, it must be at least endorsed by the authorized agent, and committed to by the responsible agent. • If the authorized agent endorses the action in front of the responsible agent who has committed to the action, the responsible agent will be expected to either execute the action (by himself or with help) or explain why not. 52 / 54 Plans for the Future - Explaining Responses • In a training simulation, it is important that the trainee understand why his proposals or actions are incorrect or inappropriate. • Work is being done on better methods of explaining and justifying the responses of the agents in to the trainee’s proposals. 53 / 54 References • Audio, Video and VR – J. Rickel et al. Towards a New Generation of Virtual Humans for Interactive Experiences . – W. Swartout et al. Towards the Holodeck: Integrating Graphics, Sound, Character and Story. • The Agent Models – J. Rickel , W. L. Johnson Animated Agents for Procedural Training in Virtual Reality: Perception, Cognition, and Motor Control. – J. Rickel Extending Virtual Humans to Support Team Training in Virtual Reality. – D. McAllester, D. Rosenblitt Systematic Nonlinear Planning. • The Negotiation Model – S Larsson, D. Traum Information state and dialogue management in the TRINDI Dialogue Move Engine Toolkit. – D. Traum, Computational Theory of Grounding in Natural Language Conversation. • More articles (and other related material) can be found at the web page of The Institute for Creative Technologies 54 / 54