Intelligent Service Agents

A note on terminology:

The term 'agent' is often used to denote a number of different things. In particular the term is used to describe both 'software agents' and 'physical agents'. Software agents are here interpreted as software systems for user assistance for applications like internet search, while physical agents denote mobile systems that perform specific tasks in a natural environment. In this paper we will *only* consider physical agents. It is, however, considered important that FIPA explicitly states its objectives with respect to both of these categories of agents.

Aim:

The Centre for Autonomous Systems at the Kungliga Tekniska Hoegskolan (Royal Institute of Technology) does research on intelligent agents. These agents are employed in service applications such as cleaning, surveillance and transportation for domestic and manufacturing markets.

Objectives:

Intelligent service agents must be able to operate in quasi-structured natural environments, where they must carry out their missions in concert with humans. This implies that the agents should have facilities for: mobility, perception, manipulation and human-agent interaction.

Limitations of existing technology:

Over the last thirty years a number of mobile robot systems has been developed, and a few of these has been employed in commercial applications. An example of such a system is the TRC HelpMate that carries out transportation tasks in a hospital environment.

Most of the existing systems are characterised by:

1. operation only in well defined environments that have controlled lighting, an pre-defined structure/layout, limited dynamics, and mobility is usually wheeled (i.e. operation is on a planar surface)
2. a rather limited number of tasks
3. must be programmed/instructed by highly trained personnel
4. limited or no manipulation.
   

To enable operation in less constrained environments it is considered lessential that robot systems be equipped with more robust and versatile perception modules. Here computational vision appears particularly promising, but it is considered essential that a variety of sensors such as GPS, gyros, sonar, etc. also be employed to ensure efficient and robust operation. A central issue is thus a standard set of 'perception modules' and facilities for sensory integration/fusion.

For large scale applications it is necessary for service agents to have a rich set of different behaviours (perception-action mappings). To combine the behaviours into operational systems it is further necessary to have composition methods that allow characterisation and reasoning about the system behaviour. The Hybrid Systems formalism from control theory is
considered promising for this.

The use of traditional modalities like mouse and keyboard is not considered an appropriate medium for user-agent interaction. To enable non-specialists to instruct the agent to carry out tasks speech and gestures appear to be the most promising modalities. It is consequently necessary to have standard facilities for speech synthesis and analysis, and facilities for interpretation of gestures. It is considered essential that the gesture interpretation can capture both spatial and temporal patterns of interaction.

For most applications it is necessary for the agent to interact with the environment. This interaction is often going to be in terms of manipulation, that can be used for pickup of items, handling of cleaners, and so forth. Most of the manipulators available today are characterised by a low payload/weight ratio, which makes them unsuited for use on a mobile platform. For many situations it is possible to complete tasks without the flexibility of a fully fledged manipulator. There is thus a need for a set of basic parts, which provide linear and rotational degrees of freedom, that can be used for flexible/modular composition of manipulation systems.

Last but not least there is a need for standard architectures and languages that enable simple composition/integration of systems from a set of basic modules. Several different paradigms have been presented but most of them do not scale to 'general' systems. It is considered essential that the paradigm enable hierarchical construction and analysis of systems. For example the Robot Schema paradigm combined with discrete event system theory appear promising. Definition of a standard implementation language is, however, considered beyond the scope of FIPA.

FIPA relationship:

To enable design and use of intelligent service agents there is, at least, a need for standardisation of the following components / systems:

1. Speech synthesis and analysis
2. Gesture interpretation for manipulation and locomotion
3. Modular components for manipulation
4. 'Standard' perception modules like vision, GPS, Sonars, gyros, ...
5. Standards for control/composition of systems (multiple paradigms)
6. Languages for instruction / task description / planning
7. Standardisation of mobile platforms
8. Standard platforms for non-wheeled locomotion (legs, tracks, ...)

There are many more issues of a more detailed nature that could be addressed. At least the items 1, 2, 3, 7, 8 and parts of 4 might be suitable for short term standardisation.

Henrik I. Christensen & Jan-Olof Eklundh
Centre for Autonomous Systems
Kungliga Tekniska Hoegskolan
S-100 44 Stockholm
Sweden
hic@vision.auc.dk and joe@bion.kth.se