Mel Slater's Presence Blog

Thoughts about research and radical new applications of virtual reality - a place to write freely without the constraints of academic publishing,and have some fun.

27 May, 2012

Inverse Presence

Inverse Presence

People use virtual reality for a purpose - like any 
kind of media the purpose of experiencing it is
to effect some change. This change might be simply 
at the level of enjoying something (entertainment), or viewing something (for the sake of understanding, design or development), or to learn something new, or to have some new kind of experience that is not available or difficult to achieve in everyday reality.

In some sense there is always a task to be realised - whether it is "enjoy yourself"or something more concrete than that such as to realise a specific set of actions. 

Now for many years there has been research on factors that contribute to 'presence'. Here by 'presence' I mean that aspect where people tend to respond to situations and events in the virtual world as if they were real: an avatar smiles at you and you smile back, or an avatar comes close to you and you feel uncomfortable and possibly step backwards, since this breaks the norms of proxemics. 

Even if only implicitly most applications of VR rely on presence occurring. For example, using VR for psychotherapy would be useless unless the patients, to some extent, respond realistically to what is depicted in the VR. So VR therapy, for example, for fear of heights would not be useful if patients did not feel some anxiety around their experience of precipices in VR.

Lydia Reeves Timmins and Matthew Lombard used the term 'inverse presence' to describe situations where something happens in reality that is perceived as if it were not real - real events (especially horrific ones) may be experienced as 'simulated' - i.e., we have all experienced moments in which we think "this is not really happening". Here I want to give a different meaning to the term "inverse presence" - to mean that we assume that presence in a VR will happen, and therefore we exploit this to get participants to achieve some particular tasks that they had never explicitly been told to do. For example, suppose the task is "get this person to smile" - then elements of the virtual environment (such as avatars) must learn to carry out actions that evoke this response. Probably here it would be quite easy - since from 'presence theory' we know that if an avatar smiles at the participant they are very likely to smile back - and so introduction of smiling avatars would probably do the trick.

Most experiment studies on presence vary factors that are thought to contribute to presence, and then see when and if presence occurs. 'Inverse presence' means - we know that presence will occur and so certain behaviours are likely to follow from this, and let's utilise it to get people to do certain specified actions.

In the recently published paper (PDF) of Jason Kastanis and myself in ACM TAP we described a quite simple example of this approach. When you interact with a virtual human character in immersive VR you tend to respond realistically. In particular several other works have shown that the rules of 'proxemics' operate - that is, if the avatar approaches too closely to you, you step backwards (the implicit rules of social, personal and intimate space seem to apply in your interactions with avatars too). Our goal was for an avatar to learn how to get the participant to go to a particular place within the virtual environment - a place some metres behind where they were initially standing. The avatar was programmed with a number of actions it could take - like move forward or back, do nothing or wave to the participant saying "come here". At first the avatar chose these actions at random, but over time it converged on the right behaviours - get the person close to the avatar and then move forward towards the person so that the person backed away. The avatar was controlled by a Reinforcement Learning agent, that received a reward when the person moved towards the target, and a loss when the person didn't do that. The RL algorithm is designed to maximise long term reward. What we found is that in circumstances when the avatar was allowed to move to intimate distance to the person, it learned how to drive them to the pre-specified place within 7 minutes. It took much longer if they could only move to personal distance, and didn't work at all if it just selected random actions (move forward or back or wave).

The purpose of VR is to get people to 'do' things (doing includes experiencing). Here we let the VR system learn how to get the person to do things by relying on their likely responses to events, as predicted by presence theory. The RL worked efficiently, but of course this was a very simple 1D problem. Nevertheless I think that the paradigm is worth pursuing with more complex scenarios.

0 Comments:

Post a Comment

<< Home