From OpenCog
Jump to: navigation, search

Embodiment: Perception and Action

Embodiment and its Importance for OpenCog Prime

Regarding the importance of embodiment for AI, expert opinions are all over the map.

  1. Some AI theorists believe that robotic embodiment is necessary for the achievement of powerful AGI (Rodney Brooks for instance is a prominent champion of this perspective)
  2. Others believe embodiment is entirely unnecessary, and one can create a human-level AGI exclusively using other mechanisms, e.g. textual conversation, interaction in formal logic, etc.
  3. We believe embodiment is extremely convenient for AGI though perhaps not strictly necessary; and that virtual-world embodiment is an important, pragmatic and scalable approach to pursue alongside physical-robot embodiment. The essay "Post-Embodied AI" elaborates this perspective in detail.

The natural synergy between advanced AI and gaming/virtual worlds has been avidly discussed for at least a decade (a milestone in this regard is John Laird's work), and is now finally becoming a practical reality. OpenCog will be integrated with a variety of virtual worlds including Multiverse and OpenSim.

Integrating Virtual Worlds with Robot Simulators

A critical question, for those who believe that embodiment is important for AGI, is: What must a world be, that an intelligent system may grow and learn within it?

And: What facilities must exist, within a mind, to mediate its interaction with other minds within this world?

Along these lines, it is arguable that current virtual world and gaming technology is not quite powerful and flexible enough to fully support virtually embodied AI. It has been argued that the way avatar actions are invoked in current virtual worlds is highly suboptimal from an AI perspective. Instead of having avatars controlled via a fixed set of animations programmed in programs like 3dsMax or specialized tools like those available for Second Life, avatars should have their joints controlled by simulated servomotors, as is done in robot simulators. Toward this end, integration of virtual world platforms with robotics simulation platforms such as Player/Gazebo would be one route.

Integration of OpenCog and Related Software with Virtual Worlds

A project to integrate OpenCog with the OpenSim virtual world is expected to start in late June 2008, and a Google Summer of Code student is also working on this during summer 2008.

Some related work has been done with the Novamente Cognition Engine:

  1. a prototype integration of the NCE with AGISim, an experimental virtual world based on the CrystalSpace game engine. A paper summarizing some of this work, which involved an integration of PLN inference, OpenCogPrime:PerceptualPatternMining, and PredicateSchematization.
  1. the Novamente Pet Brain, a specialized incarnation of the NCE, has been used (on an experimental, not yet productized basis) to control intelligent virtual pets in the Second Life and Multiverse virtual worlds. Some papers describing aspects of this work are available.

Perception Processing in OpenCog Prime

Some detailed thoughts on the processing of perceptual data, and the generation of actions, in OpenCog Prime are at OpenCogPrime:PerceptionAction

Goal Oriented Cognition and World Interaction in OpenCog Prime

A detailed account of the "perception-cognition-action loop" via which OpenCog Prime is intended to interact with a world (physical or virtual) is at OpenCogPrime:GoalDrivenCognition.