Ruth Aylett ----------- TITLE: Exploring behaviourally-driven agents with Virtual Teletubbies --------------------------------------------------------------------- AUTHOR: Ruth Aylett, Centre for Virtual Environments, University of Salford ADDRESS: CVE, Business House, University of Salford, Salford, M5 4WT, UK 1. Cooperating robot work at Salford At Salford, work has been carried out over the last five years into multiple cooperating robots [1]. A novel behavioural architecture, the Behavioural Synthesis Architecture (BSA) was developed [2] consisting of behaviour patterns, behaviour packets and behaviour scripts. We keep the term 'behaviour' on its own to refer to the overall emergent behaviour of a roboot. In the BSA, emergent behaviour is prouced by currently active behaviour patterns, each of which is represented as a pair of functions. One function maps incoming sensor stimulus to actuator response, whicle the other maps the same sensor stimulus to utility, which is a measure of how important the response is for that stimulus. A synthesis mechanism then combines all active patterns as weighted by their utilities to produce an emergent response for that actuator. Note that this architecture differs from subsumption since in that case, the stimulus-response functions were combined into a network such that only one function had complete control of the actuators at any given moment and emergent behaviour was produced by time-slicing. In the BSA all active patterns contribute to emergent behaviour continously via the synthesis mechanism. For conceptual convenience, behaviour patterns in the BSA were divided into four categories. These were: 1.SELF: behaviours to do with self-preservation, like recharging a low battery 2.SPECIES: behaviours concerned with interaction between robots 3.ENVIRONMENT: behaviours such as obstacle avoidance, concerned with moving around the environment 4.UNIVERSE: Behaviour such as navigating to a particular beacon concerned with overall task achievement A known problem with behavioural architectures is that interactions between behaviour patterns may produce undesirable effects. For example, patterns designed to produce obstacle avoidance might not be very useful for an agent that was trying to sit on a chair. To overcome this problem, the BSA makes use of sets of active behaviour patterns called behaviour packets. Each packet consists of a sensory pre-condition (for example, an agent'notices' a particular object)., the names of behaviour patterns to be set 'active', and a sensory post-condition at which the packet ends. All that is then required is a mechanism for switching between behaviour packets according to the context, and this is supplied by a behaviour script. This is a collection of packets, in our robot work chained together sequentially, reflecting the needs of robot task-based environments. However there is no reason in principle why all packets should not be available for activation at any given time rather like rules in a production system or indeed why more complex schemes for packet activation should not be tried. 2. Virtual Teletubbies It was decided to reapply this robot architecture to agents in a virtual environment and the scenario chosen was that of virtual Teletubbies [3] - characters from a well-known television series for very small schildren in the UK. The requirements for the system were derived from the TV programme since an aim was to faithfully reproduce the 'look and feel'. Note however that a TV programme is necessarily linear, with a beginning, middle and end, following a story (albeit a simple one). A VE is much more open, and although work such as that in the OZ project has investigated VEs as interactive drama, it was decided to concentrate initially on producing authentic behaviour which would interest the user without forming part of an overall 'story'. Anaysis suggested that the following behaviours were required: Basic wandering by teletubbies, with contour-following and obstacle avoidance Ability to 'notice' objects in the environment, including each other Ability to head for dome and consume Teletubby toast or custard when hungry Ability to head for windmill when it rotates Teletubby group behaviours such as hugging In addition, the Noo-noo (vacuum cleaner) can be thought of as a fifth Teletubby with behaviours required to notice 'mess' and clear it up while the windmill and the toaster (for making Teletubby toast) can also be thought of as agents with limited capabilities. Behaviour patterns and packets were implemented as described above, using the data from virtual sensors attached to a Teletubby as stimuli and the Teletubby graphical representations as actuators. Advantage was taken of the extra information available within the VE as against a real-world domain so that where a robot might just get an infra-red reading from a nearby object, a Teletubby gets information about what the object is. A more sophisticated scheme was adopted at the behaviour script level than that which had been implemented for the BSA in robots. The behaviour script level is concerned with sequencing of behaviour, and while this is straightforward for a single robot task, the open-ended behaviour of Teletubbies in a VE effectively requires an endless behaviour script which is constructed dynamically by interaction with the environment, including other Teletubbies. In order to achieve this, four queues were constructed, one for each of the conceptual categories SELF, SPECIES, ENVIRONMENT and UNIVERSE referred to in section 2. The entries in this queue consist of groups containing one or more behaviour packets, effectively sub-scripts, each with an attached priority. The subscript with the highest priority is then selected for packet execution. For example, an agent property is hunger, which is increased over time. A very hungry Teletubby should ignore any other stimuli and execute a script which takes him or her to the Dome at once to feed up on custard or toast. The drives that were implemented to drive subscripts were hunger, curiousity and tiredness. A rudimentary physics was created for the world by modelling gravity in a simple way through the equation: f= m * a where f is force, m is mass and a is acceleration. Teletubbies then output force which is transformed into acceleration and was used to make Teletubbies move faster as they go downhill. It also prevented the user avatar from leaving the ground - if a user tries to do so they effectively execute a jump as gravity brings them back to the ground. In the same way, if a Teletubby is lifted off the ground and dropped, it will land and bounce. This emerges from the physics and does not need to be explicitly programmed. 3. Issues arising from our work The architecture that was developed was successful in achieving a reasonable subset of the desired behaviours listed above. Cooperative behaviours such as hugging have not been tackled so far however since it is not yet clear how to achieve this sort of physical interaction in a principled manner as distinct from superimposing an animation. The usual (for behavioural systems) difficulties in fine-tuning behaviour patterns so as to produce useful and interesting results were encountered. An area we have identified for future work is how behavioural systems like this can be used to produce dramatic interest or narrative streams which would involve the user. We are interested in extending the idea of emergence, which is used in robotics to produce physically coherent behavior, to a higher level of emotionally or socially coherent behaviour. Finally, our mobile robot work shows that it is possible to interface our behavioural architecture to a predictive planning system through behaviour scripts. In the robot case, this produces a system driven by user goals, but it would be interesting to allow the Teletubby drives to produce goals which could then be fed through such a system in order to structure - though not determine - their low-level behaviour. References: [1] http://www.salford.ac.uk/eee/research/mobrobs/fandg.html [2] "Many hands make light work? An Investigation into behaviourally controlled co-operant autonomous mobile robots" D.P.Barnes, R.A.Ghanea-Hercock; R.S.Aylett & A.M.Coddington (1997) . Proceedings, 1st International Conference on Autonomous Agents, Marina del rey, Feb 1997 obtainable from http://www.salford.ac.uk/iti/projects/MACTA/macta-papers.html [3] 'Virtual Teletubbies: reapplying a robot architecture to virtual agents' Aylett, R.S; Horrobin, A.J; O'Hare, J; Osman, A.A. & Polshaw, M.M.T. http://www.salford.ac.uk/iti/rsa/teletub.html Ruth Aylett email: R.S.Aylett@iti.salford.ac.uk phone: 44-161-295-2912 Home page: http://www.salford.ac.uk/iti/rsa/aylett.html fax: 44-161-295-2925 Centre for Virtual Environments, Business House, University of Salford, Salford, M5 4WT, UK 'Life is beautiful.'