|AAAI Fall Symposium on|
Anchoring Symbols to Sensor Data |
in Single and Multiple Robot Systems
The focus of this Symposium was the connection between symbol- and signal-level representations of physical objects in autonomous robotic systems. We call "anchoring" the process of creating, and maintaining in time, this connection.
Anchoring must necessarily occur in any physically embedded system that comprises a symbolic reasoning component. A typical example is the problem of connecting, inside an autonomous robot, the symbol used by a planner to refer to a particular box, say `box-21', to the data that correspond to that box in the sensori-motoric system. This connection must be dynamic since the same symbol must be associated to new entities in the perceptual stream in order to track the object over time or to re-acquire it at a later moment. Anchoring must also occur in a multiple robot system, since the robots must agree about the meaning of the symbols used to refer to perceived objects in the environment.
Solutions to the anchoring problem are currently developed on a system by system basis for restricted domains. The aim of this Symposium was to create an inter-disciplinary community interested in the development of general theories of anchoring. Having such a theory will greatly advance our ability to build intelligent embedded systems, and to transfer techniques and results across different systems.
Attended by approximately 30 participants, the Symposium was successful in achieving its aim. On the first day, each attendant was given two minutes to introduce their group and their research in a highly dynamic "rump session". The participants especially appreciated this session since many of them came from different communities and did not know each other. The remainder of the Symposium was largely dedicated to the presentation of full-length papers, two invited talks, and even a virtual presentation, which covered different facets of the anchoring problem. Topics included the description of deployed robotic systems that use anchoring; the bottom-up extraction of symbols from sensor data; the anchoring of linguistic terms in human-robot or robot-robot communication; and related issues like visual attention and conceptual spaces. Presentations differed significantly in the nature of the entities to be anchored: some author focused on symbols that denote individuals, while others focused on symbols that denote categories, actions, or events.
To further encourage creative interaction between the participants, the Symposium included three panel discussions. Two short panels were devoted to the discussion of the relation between anchoring and other problems, including symbol grounding, pattern recognition, and tracking. A longer panel was organized at the end of the Symposium, and had the ambition to set up a research agenda for the anchoring problem. A crucial observation was that although the anchoring problem can easily become extraordinarily complex, we nonetheless have to deal with it if we want to build robots that manipulate their environment and interact with their users. There was general agreement that a first priority is to identify the right level of abstraction for a general theory of anchoring: this should be specific enough to lead to tractable solutions, but still general enough so that results and techniques can be ported across different robots and domains.
Besides the organized panels, a lot of spontaneous discussions arose during and after the presentations. Interestingly, a few recurrent themes emerged from these discussions, which appear to point to some fundamental issues for the anchoring problem. For instance, it was noted that in most presented systems anchoring relies on the existence of an internal representation for an object that is somehow intermediate between the symbol and the perceptual data. This intermediate representation plays a pivotal role in connecting different sub-systems (e.g.: perception, reasoning, motion) which all need to refer to the same physical object. A second observation was that most presenters dealt with anchoring either bottom-up (given some perception, associate it to a new or a pre-existing symbol) or top-down (given a symbol, find the corresponding data in the perceptual stream). It was felt that a general solution to anchoring should integrate both directions. Finally, a question often popped up of whether we should focus on symbols that correspond to human-like categories or symbols that correspond to categories closer to the robot's senses. Although the latter might be a easier starting point, we will eventually need the former in order to have our robots interact with humans in everyday life.
[ Symposium Home ]