On AI Emobodiment in VWs

Embodiment is an old concept in general purpose AI, creating a dividing line between researchers, with supporters of good old fashioned AI stressing embodiment in a limited sense, while people like Rodney Brooks and others insist that embodiment is the key to a general purpose AI.

Current embodiment research focus mainly -with a few visionary exceptions- on robotics, with big names in the industry like Honda trying to emphasize humanoid robots. However there has been a growing number of AI researchers pushing for VWs embodiment. The concept can be traced back as far as 1968 when SHRDLU was invented and later by 1970 a virtual environment (SHRDLU's "world") was added.

SHRDLU's world with color rendering (Univ. of Utah)

The current state of VWs provide a much richer experience, and much more flexibility. Moreover, with the vast numbers of people connecting, playing and interacting in virtual worlds, it provides coupling between embodied AI and humans to allow the AI to learn about itself, the world, and others.

The coupling between an avatar and an automated general purpose embodied AI is not the only possible coupling. VWs alow for a coupling between physical properties and an embodied AI, where the automated avatar interacts with objects that possess some physical properties allowing it to formulate hyposis about the nature of the physica environment. For example in SL the sun rises and sets every 4 hours, gravity (things fall down and float up), persistence of objects, solidity of objects, ..etc.

Yet another possible coupling is between the AI avatars themselves allowing them to share their newly acquired knowledge and representation from interactions with the universe and the other avatars, creating a 'wisdom of the crowd' scinario.

In my oppinion, while embodiment should not be a goal by itself, VW embodiment provide for a contextual learning experience. Where objects can be accessed in the context of the environment, the collective knowledge and the interactions.