Notes on "Embodied AI as Science: Models of Embodied Cognition, Embodied Models of Cognition, or Both?"

"Embodied AI as Science: Models of Embodied Cognition, Embodied Models of Cognition, or Both?" is by Tom Ziemke and it was published in 2004.

This paper is available at Google Books and SpringerLink.

One of the reasons that I am reading this paper is to clarify my definition of embodied AI. Unfortunately, Ziemke, 2004, notes that such a definition is currently fuzzy.

1.0 Introduction

Ziemke holds that embodied AI research needs to take further the concept of embodiment being important to cognition.

This is quite a valid point based on the book I read recently, "How the body affects the way we think" by Pfeifer and Bongard, 2007.

Symbolic grounding has been the core focus but researchers are beginning to hold that this focus is too narrow.

2.0 Background: What is Embodied AI anyway?

2.1 Motivation

Embodied AI research treats the concept of embodiment with less depth than related fields.

Researchers tend to define embodied AI in terms of what it is not, too much.

2.2 Embodied AI: Science versus Engineering

Engineering is concerned with the design of artefacts and science is concerned with the understanding of natural systems.

Physical models are important than simulations in engineering. However, this is not necessarily the same in science.

3.0 Notions of Embodiment

Nunez, 2004, distinguishes embodiment as follows. Trivial embodiment is the view that intelligence relates to biological processes. Material embodiment is the view that low-level intelligence requires a body. Full embodiment is the view that all forms of intelligence require a body.

Clark, 1999, distinguishes embodiment as follows. Simple embodiment is the view that some types of intelligence require a body. Radical embodiment is the view that all intelligence requires a body.

Wilson, 2002, discusses another description of embodiment, which has six views. Firstly, intelligence is situated. Secondly, intelligence is time pressured. Thirdly, intelligence is for action control. Fourthly, humans off-load intelligence work onto the environment. Fifthly, the environment is part of the intelligence system. Sixthly, off-line intelligence is body based.

The sixth point requires further elaboration.

The following are additional views on embodied intelligence.

The structural coupling view holds that coupling between the agent and the environment does not necessarily require a physical body.

The historical embodiment view holds that embodiment is a result of a history of coupling and adaptation towards an ecological niche and a physical body is not necessary.

The physical embodiment view holds that a physical body is necessary.

The organismoid embodiment view holds that a physical body is required and that morphology is critical. There are two more views under the organismoid embodiment view. Firstly, a body mediates between internal processes and the environment. This is similar to the material embodiment view. Secondly, intelligence depends on the sharing of the control system between the sensorimotor and abstract intelligence processes. This is similar to the full embodiment view.

The organismic embodiment view holds that some aspects of intelligence depend on autopoiesis. Autopoiesis means self-creating, maintaining and organising of living bodies.

4.0 Discussion: Implications for Embodied AI as Science

Questioning methodology is part of the scientific process.

The classical or cognitive approach has been to use computer programs as models of intelligence.

The use of physical robots is not the only defining point for embodied AI.

The use of embodiment concepts as a basis for theories of intelligence is not the only defining point for embodied AI.

The use of full embodiment concepts is not the only defining point for embodied AI. Full embodiment is the view that all intelligence requires a body.

The combination of physical robots and of theories of intelligence based on embodiment concepts is also insufficient.

The simple embodiment view includes too much of the field and the full embodiment view includes too little of the field.

Although a single framework is not required, further clarification would benefit the field.

My Thoughts

Something about Ziemke's writing makes it very hard to read. The sentences are long and the paper conveys lots of information.

I am not sure what I got from this paper. I guess I got a bit more of a grasp of the state of embodied AI research.

I am realising now that describing the two approaches to AI research as I see them is justifiably difficult. It is hard describing classical AI, current AI and especially embodied AI.

I like how Dreyfus, 1992, describes what embodiment is and why it important. Philosophers have a way with asking questions and questioning answers.