Sergeiy Sandler has posted an interesting paper describing an approach to understanding linguistic content that seems applicable to AGI and robotics. The paper is titled "Reenactment: An embodied cognition approach to meaning and linguistic content" (PDF format) and is due to appear in the journal, Phenomenology and the Cognitive Sciences. He describes the goal of his paper like this:
My aim was to provide support to a vision of language and cognition that is entirely action-based, where action and interaction (performed or perceived) are the main currency in which human minds conduct their business. This vision also animates at least some of the approaches within EC. Following and developing it can also, so I believe, improve and deepen our understanding of language and meaning
Embodied Cognition is basically the idea that intelligence is closely related to the physical form, and even that the environment forms part of the cognitive system. For example EC would argue that making an intelligent robot should be easier than making the traditional brain-in-box type of AI because the robot provides a physical body equipped with sensors and actuators that interact with the real world. (If this sounds familiar, it's similar to what Rodney Brooks and Hans Moravec have argued with respect to robots). EC further involves the idea of learning and understanding through reenactment, an idea that has been supported by the discovery of mirror neurons. In his paper, Sandler argues that reenacting or simulating the actions described in utterances provides an account of linguistic meaning compatible with Embodied Cognition.