Theoretical Artificial Intelligence Foundations: An Analysis of Rodney Brook’s Approach to AI Applied

to the MIT Kismet Humanoid Bot

Juan Antonio Martín Checa

Computer Science Department

Campus Teatinos University of Malaga 29071 Málaga, Spain

jamcheca@telefonica.net

Abstract. Despite the fact that the strongly adopted traditional approach of decompossing intelligent systems into independent information processing units interfacing each other via representations was the common rule in the late 80’s, MIT Professor Rodney Brooke focused his work on a brand new philosophy, according to which intelligent systems are to be divided into independent and parallel activity producers which interfere directly to the real world through perception and action. In this paper we will argue the validity and significance of Brook’s innovative contribution. Also, we will study the compliance of the MIT Kismet Humanoid Robot regarding Brook’s innovative approach.

Keywords: Abstraction, Action, Activity Producer, Artificial Intelligence, Central System, Humanoid Robot, Incremental Intelligence, Inhibition, Perception, Peripheral System, Subsumption Architecture, Suppression, Representation.

1 Introduction

The relatively short story of AI is full of ups and downs derived from a generalized misunderstanding of what AI was or not capable of achieving at the moment. High expectations were soon put on the shoulders of a scientific field still on its very early stage [1]. Companies from all over the world hired engineers in order to keep on track with the prevalent AI-era in an attempt to avoid falling behind increasing competitors.

1.1 Traditional approaches: the issue of abstraction

One of the main issues for AI researchers in the past was figuring out how to divide the system into smaller manageable pieces so that the global problem could be solved. The basic commonly accepted idea was to partition the system into two components: the AI component, focus of their work, and the non-AI component, out of their responsibility [2].

During the late 60’s and early 70’s many AI researchers explored a new approach based on a simple and uniform semantic: the so-called blocks world. The key was to explicitly and completely represent the state of the world. According to Rodney Brook, emeritus professor of robotics at MIT [5], the main problem faced while using this approach is the fact that researchers have to abstract away most of the details in order to express the initial problem in terms of atomic concepts. On Brook’s words, “abstraction is the essence of intelligence”, so by having the experimenters abstracting at that level, they are leaving little for the intelligent system to do. This is, “a truly intelligent program would [...] perform the abstraction and solve the problem.” [2].

2 Incremental intelligence

The process of implementing an intelligent system is anything but obvious. As explained by Brook, “[...] is is necessary to decompose a complex system into parts, build the parts, then interface them into a complex system.” [2]. Two valid engineering approaches for Brook at this point are either decomposing the system by functions or by activities.

2.1 Decomposition by function

This approach structures any intelligent system into three main components: a central system, perceptual modules (which sense the world, inputs), and action modules (responsible for taking actions, outputs), being both inputs and outputs expressed as symbolic descriptions. On its side, the central system must be subdivided into smaller components. Brook criticizes this approach arguing that researchers working on individual modules define each own interfaces based on their own specific needs, making it extremely hard the process of final integration of all the modules into a complete system [2].

2.2 Decomposition by activity: subsumption architecture

Since the days of Alan Turing, AI has long focused on symbolic computational approaches to creating intelligent machines [6]. Brooks focused instead on biologically-inspired robotic architectures capable of addressing basic perceptual and sensorimotor tasks.

Unlike the previous approach, the decomposition by activity partitions the global system into a number of layers called activity (behavior) producing subsystems, each of them capable of connecting sensing to action in an individual fashion. The key here is that both perception and action are not centralized but distributed throughout the global system [2]. The sum of all the layers put together and organized into a predefined hierarchy is given the name of subsumption architecture [7].

As opposed to more traditional AI approaches, subsumption architecture is founded on a bottom-up design [7]. The main advantage of adhering to this approach

is the fact that it makes it easier the process of implementing an intelligent system following a a-priori-unlimited number of incremental steps allowing the system to grow up in intelligence as much as needed and possible at the moment. Finally. each of the layers must be extensively debugged in the real word before a new layer is added to the system [2].

3 Subsumption architecture Vs mechanistic approaches

When analyzing Brook’s approach of subsumption architecture, one may think it shares aspects with mechanistic approaches such as connectionism, neural networks, production rules, and blackboard systems. However, Brook steps aside, criticising them for a number of reasons.

Regarding connectionism, the nodes used are not only uniform but also extensively interconnected, while in subsumption architecture all nodes are unique finite state machines with low connections between layers [2]. Furthermore, there seem to be a problem with the explicit distributed representations that connectionists expect to spontaneosly arise from their networks. As stated by Brook, “[...] we believe representations are not necessary [...]” [2].

As of neural networks, these are claimed to have some biological significance. In any case, Brook confront this claim arguing that the number of modeled connections is by far much lower than that founded in human neurons.

In relation to production rules, these are based on the principle that when the right conditions can be found in the environment, a certain action will be performed. The same occurs with the subsumption architecture with the difference that while in production rules every single rule is managed by a central database, in subsumption architecture, layers run in parallel, and extracted aspects of the world directly trigger certain behaviors on a specific layer.

Finally, in blackboard systems the producer of a certain piece of knowledge is hidden to its final consumer. Subsumption architecture makes such connections explicit and permanent [2].

4 Analysis of Brook’s expectations on MIT Kismet humanoid bot

As described in [3], “Kismet is an expressive robotic creature with perceptual and motor modalities tailored to natural human communication channels.” In addition, Kismet was implemented following the approach of subsumption architecture, where “competencies are progressively modified, adapted, and built upon to produce more sophisticated, diverse, or new abilities” [3]. But, does Kismet really meet Brook’s expectations for the so-called Creatures? Next, we will analyze to what degree Kismet fits into Brook’s concept of Creature.

On his paper Intelligence without representation, Brook enumerates some of the necessary characteristics that should be present on any Creature. We will present them all and study whether or not they can be found on Kismet.

The first requirement is that a Creature should be capable of coping appropriately and in a timely fashion with changes in its dynamic environment. Kismet’s low-level feature extraction system is based on fast routines, so that it can respond rapidly to changes in the environment. Indeed, when Kismet is exposed to fast approaching objects, its looming reflex is triggered causing the robot to quickly withdraw or even escape, emulating natural human behaviors[3].

The second requirement is that a Creature´s behavior should be consistent and robust, avoiding the Creature to collapse. Kismet’s behavior system is structured into a coherent structure in which each possible behavior is viewed as a “self-interested goal-directed entity that competes with other behaviors to establish the current task” [3]. The key is the existence of an arbitration mechanism responsible for determining which behavior should be activated at any time, and for how long, based on relevancy, coherency, persistence, and opportunism. All these aspects can be found in Maslow’s Hierarchy of Needs (see Fig. 1) according to which, human beings must satisfy all lower-level needs before satisfying any specific need [4].

Fig. 1. The Physiological, Safety, Love/Belonging, Esteem, and Self-actualization levels of Maslow’s Hierarchy of Needs.

The third requirement is that a Creature should be able to pursue different goals depending on the external circumstances. Kismet’s motivation system consists of a number of drives (basic needs) and emotions. Depending on the level in which the drives are satisfied, Kismet is capable of modifying the importance associated to different goals. For instance, when the robot feels it’s in danger, the predominant goal is to escape. Once the robot feels safe, the goal of approaching and interacting with humans becomes stronger while the goal of safety stays on background.

The final requirement is that a Creature should have some purpose in being. This is, it should have a mission in the world. This is directly related to the self-actualization level of Maslow’s hierarchy of needs. As of Kismet, it has the ability of engaging and maintaining a “natural and intuitive social interaction with a human” [3]. This is Kismet’s reason for ‘existing’.

References

1. Artificial Intelligence - Wikipedia, http://en.wikipedia.org/wiki/Artificial_intelligence 2. Brooks, R.A.: Intelligence Without Representation. Artificial Intelligence. 47, 139-159

(1991) 3. Kismet, http://www.ai.mit.edu/projects/humanoid-robotics-group/kismet/kismet.html 4. Maslow’s Hierarchy of Needs,

http://en.wikipedia.org/wiki/Maslow%27s_hierarchy_of_needs 5. Rodney Brooks - Home, http://people.csail.mit.edu/brooks/ 6. Rodney Brooks - Wikipedia, http://en.wikipedia.org/wiki/Rodney_Brooks 7. Subsumption Arquitecture, http://en.wikipedia.org/wiki/Subsumption_architecture