1. Why Autonomous Virtual Characters?

Virtual characters, also known as virtual actors, have become very popular in the last ten years, mainly through three-dimensional (3D) movies and games. In movies, they now have very realistic physical and emotional characteristics, including their facial expressions, hair, clothes, and motions. But everything about these characters was rendered in an offline mode; they have no mind at all. By contrast, games such as Quake and Tomb Raider create virtual worlds where the characters dynamically interact with the users. Moreover, crowds of virtual characters have appeared not only in movies but also in games. Such characters are autonomous.

Autonomous virtual characters (AVCs) are not just for movies and games. In the future, they can be at the heart of the simulation of activities for a variety of purposes, including education and training, treatment of psychological problems, and emergency preparedness. Imagine the following scenarios:

• A user is being trained to perform some complex task, such as repairing a copy machine. He uses an interactive user manual, where an autonomous character plays an expert, showing him how to proceed. At every stage, the user is able to see what to do next, even when mistakes are made.
  
  
• A therapist is helping a patient overcome a fear of public speaking. To overcome this fear, the patient has to perform while immersed in a virtual environment consisting of a seminar room and a virtual audience, which can react to the user in an autonomous way. The therapist can choose the type of virtual audience (for instance, one that is aggressive or sexist) that will result in a more effective treatment for the patient.
  
  
• A user is learning basic life support (BLS) procedures. She is immersed in a virtual setting, and discovers a victim lying on the ground. She has to give him BLS through her proxy, a virtual assistant (VA). The user navigates the scene, assesses the situation, and makes decisions by issuing natural voice commands. The VA waits for commands and executes the actions. If the user's commands are correct, the victim recovers. In cases where the user provides incorrect commands, the VA may refuse to do harm to the victim; in such situations, the VA may prompt the user for retrial, or may suggest an alternative possibility.

Autonomy is generally the quality or state of being self-governing. Rather than acting from a script, an AVC is aware of its changing virtual environment, making its own decisions in real time in a coherent and effective way. An AVC should appear to be spontaneous and unpredictable, making the audience believe that the character is really alive and has its own will.

To be autonomous, an AVC must be able to perceive its environment and decide what to do to reach an intended goal. The decisions are then transformed into motor control actions, which are animated so that the behavior is believable. Therefore, an AVC's behavior consists of always repeating the following sequence: perception of the environment, action selection, and reaction.

The problem with designing AVCs is determining how to decide on the appropriate actions at each point in time, to work toward the satisfaction of the current goal, which represents the AVC's most urgent need. At the same time, there is a need to pay attention to the demands and opportunities coming from the environment, without neglecting, in the long term, the satisfaction of other active needs.

There are four properties that determine how AVCs make their decisions: perception, adaptation and intelligence, memory, and emotions.

Perception. Perception of the elements in the environment is essential for AVCs, as it gives them an awareness of what is changing. An AVC continuously modifies its environment, which, in turn, influences its perceptions. Therefore, sensorial information drastically influences AVC behavior. This means that we cannot build believable AVCs without considering the way they perceive the world and each other. Imagine if, in the real world, everyone could see the whole world, and could hear everything. This is the situation in virtual reality, so the issue for AVCs is how to filter this information correctly. To realize believable perception, AVCs should have sensors that simulate the functionality of their organic counterparts, mainly for vision, audition, and tactile sensation. These sensors should be used as a basis for implementing everyday behaviors, such as visually directed locomotion, responses to sounds and utterances, and the handling of objects.

Adaptation and intelligence. Adaptation and intelligence define how the character is capable of reasoning about what it perceives, especially when unpredictable events happen. An AVC should constantly choose the best action so that it can survive in its environment and accomplish its goals. As the environment changes, the AVC should be able to react dynamically to new elements, so its beliefs and goals may evolve over time. An AVC determines its next action by reasoning about what it knows to be true at a specific time. Its knowledge is decomposed into its beliefs and internal states, goals, and plans, which specify a sequence of actions required to achieve a specific goal. When simulating large groups or communities of AVCs, it is possible to use bottom-up solutions that use artificial life techniques, rather than top-down, plan-based approaches, such as those that are common in artificial intelligence. This allows new, unplanned behaviors to emerge.

Memory. It is necessary for an AVC to have a memory so that similar behaviors can be selected when predictable elements reappear. Memory plays an important role in the modelling of autonomy, as actions are often decided based on memories. But imagine an AVC in a room containing 100 different objects. Which objects can be considered memorized by the virtual character? It is tempting to decide that whenever an object is seen by the AVC, it should be stored in its memory. But if you consider humans, nobody is able to remember every single object in a room. Therefore, the memory of a realistic AVC should not be perfect either.

Emotions. The believability of an AVC is made possible by the emergence of emotions clearly expressed at the right moment [1]. An emotion is an emotive reaction to a perception that induces a character to assume a physical response, facial expression, or gesture, or select a specific behavior. The apparent emotions of an AVC and the way it reacts are what give it the appearance of a living being with needs and desires. Without them, an actor would just look like an automaton. Apart from making them appear more realistic, AVCs' visible emotions can provide designers with a direct way of influencing the user's emotional state.

3. The Impact of Research On Autonomous Virtual Characters

The four properties above are very important in creating believable autonomous characters. Modeling these properties accurately in real time requires research efforts from various branches of computer science. We emphasize a few of them below.

Behavior Planning. Behavior planning involves the selection of appropriate actions for the AVC to execute. These decisions should reflect the individual characteristics of the AVC, including its intelligence, its motivations, and its social behavior. Besides being individual, action selection architectures for AVCs should be both reactive and proactive to be efficient in real-time. The transitions between reactions and planning should be rapid and continuous in order to elicit coherent and appropriate behaviors in changed or unexpected situations. The design of a behavior planner satisfying these criteria is a research challenge.

Virtual Sensor Design. It is tempting to simulate perception by directly retrieving the location of each perceived object straight from the environment. For realistic AVCs, this is not appropriate; instead, virtual sensors are used to simulate an AVC's perception of its virtual environment. What is important is the functionality of a sensor and how it filters the information flow from the environment.

It is not necessary or efficient to model sensors with biological accuracy. Therefore, virtual eyes may be represented by a Z-buffered color image representing a character's vision. A virtual nose, or other tactile point-like sensors, may be represented by a simple function evaluating the global force field at the sensor's location. The virtual ear of a character may be represented by a function returning the ongoing sound events. With these virtual sensors, AVCs should be able to perceive the virtual world in a way that is very similar to the way they would perceive the real one.

Research advances introduce proprioception into a unified perception concept for an AVC in a situated virtual world. The motivation is to reach persistency and to obtain a cognitive map of the perceived virtual world. It is also possible to integrate a perception approach by including the faculty of prediction, for example, the orientation of the AVC’s attention.

In addition to perceiving the virtual world, some AVCs have to be aware of certain events and characteristics of the real world. It takes real devices, such as cameras, microphones, and haptic devices, to capture this information and bring it to the virtual characters. The information they provide must be integrated with that of virtual sensors.

Modeling Emotions. To allow AVCs to respond emotionally to a situation, they could be equipped with a computational model of emotional behavior. Emotionally related behavior, such as facial expressions and posture, can be coupled with this computational model, which can be used to influence their actions. The development of a good computational model is a challenge.

4. Ethical concerns

There are ethical concerns regarding the use of AVCs. One such concern involves decisions made by real people that are based on the advice of autonomous characters. Autonomy means that the AVC makes decisions based on his or her understanding of the environment and of the rules for the surrounding world. In a simulated world, how can we be sure that this information corresponds to reality? As with any computer program, AVCs are not immune to bugs or tampering. Their advice on critical matters should always be validated.

Another ethical concern is the fact that an autonomous virtual human may be indiscernible from a real existing person. For example, a terrorist group could create a TV spot showing democratic leaders promoting nondemocratic values. To avoid misleading and manipulating the public, we will need to use technology, such as watermarking, to reliably indicate to the viewer that the human is virtual.

Some autonomous characters promote violence, terrorism, abuse, or crime, in the context of games or other interactive situations. If their behavior is realistic, they are likely to exert a strong negative influence, even when they are known to be virtual. New laws and regulations will have to be developed in this area.

5. Scalability and mobility

With the advent of wearable devices, advanced PDAs, and smartphones, AVCs can be with people all the time to guide and help them. Moreover, with light see-through head-mounted displays, it becomes easy to add virtual characters to a real scene. Showing someone how to use an electronic device or how to find his or her way are some possible applications.

6. Possibilities and challenges

AVCs will be essential in many applications of the future. They will be our playmates, teachers, therapists, and pets. Because of their logic and memory, autonomous characters will bring skills and abilities that complement those of humans, rather than replace them. The next generation--children and young adults--is very open to virtual communities and e-learning, and is expected to have no problems interacting with virtual humans.

How far are we from such a situation? Current AVCs are becoming more realistic in terms of their appearance and animation, and they are able to perceive the virtual world, and the people living in that world. They may act based on their perception in an autonomous manner. However, their intelligence is constrained and limited.

In the near future, we may expect to have AVCs that are able to learn or understand a few situations, due to the development of new methods of artificial intelligence. However, a great deal of research effort is still needed to reach the point at which AVCs can behave autonomously and interact naturally, like real creatures. This is especially the case for simulating autonomous virtual humans. True imitation of the full complexity of human behavior may not be accomplished even by the end of this century.