1 The GeckoSystems Architecture for Robot Intelligence and Control

The human biological machine, whether it be God's creation or some product of evolutionary processes, should be considered a marvel of composite complex processes.   And regardless of the genesis stance that one takes, it is impossible to deny the near perfection in terms of utility and function provided by the human "design" (for lack of a better word).   Those who seek to produce systems mimicking the functionality of the human body and mind without recourse to the ideal model provided by nature may well be ignoring the ultimate in time-tested and elegant solutions.   GeckoSystems, Inc., in understanding the synergy of drawing from both biological and traditional engineering approaches, has developed a hybrid paradigm for robot intelligence and autonomous control.

GeckoSystems' approach is a modular, distributed and hierarchical robot control scheme comprised of two abstract components: an upstream component where deliberative and semi-real-time processing occurs and a downstream component where real-time reactive-level control is achieved.   This architecture is currently implemented in all of our platforms including the CareBot™, SecurityBot™, and SuperVacBot™.   In this article, we describe both components in detail, along with inherent advantages of our architecture.   Additionally, we address several potential drawbacks of our architecture over flat architectures.   We will use biological metaphors to help the reader understand the organization behind the architecture, drawing from the demonstrable robustness of both human and animal biology.   Furthermore, we have refrained from identifying specific hardware solutions to further illustrate portability of the software.

2 Upstream Cognitive Lobes

The upstream area of the GeckoSystems architecture is involved in higher-level, semi-real time and deliberative processing.   This region is likened to the human brain's cerebrum, which is a collection of tightly-coupled, specialized processing units called lobes including the temporal, occipital, parietal and frontal lobes.   Together, these lobes define the human ability to think, contemplate and plan, also called cognitive processing.

2.1 GeckoNav™

Each lobe of the human brain contains large areas called association areas, where it is believed that problem solving is performed and memory is stored.   Association areas, along with a neuron bundle called the corpus callosum, serve to interconnect adjacent lobes.   Because GeckoNav is involved in both centralized control and environmental mapping as well as in resolving communication between junior upstream components such as GeckoChat and GeckoVision, we will compare GeckoNav with this conjoined mass of association areas.   In effect, GeckoNav is the supervisor for both the upstream and downstream systems and subsystems.

GeckoNav is the central control unit of the upstream component, being intimately involved in the crucial task of navigation.   Other high-level units, such as GeckoChat, are subservient to GeckoNav in that navigation and control issues are directed through it.

The GeckoSystems Architecture

GeckoNav handles subsumption (prioritization in context) over conflicting internal representations and external signals.   Additionally, GeckoNav provides the only interface to the downstream component, thus isolating and protecting other upstream nodes from low-level control issues.   For example, when driving in dense traffic, one generally has difficulty conversing.   Thus, for safety reasons, the human brain understands the importance of not engaging in peripheral distractions.

GeckoNav's chief role is in providing top-level navigation for the interfaced platform.   By using a subsumptive set of proprietary algorithms for mapping, path planning, obstacle avoidance and environment exploration, GeckoNav, in tandem with the downstream components, successfully and demonstrably (see video demos) solves the problem of autonomous, real-time robot navigation in dynamic domains.

2.2 GeckoChat™ (GeckoTempora)

The human cerebrum is divided into several regions called lobes, which are further subdivided into areas according to specific functionality.   The lobe that controls the majority of auditory processing is called the temporal lobe, which also contains Wernicke's area, one of the two areas associated with speech. GeckoTempora is the term we give for our speech synthesis and recognition approach, with the software product GeckoChat being the manifestation.

GeckoChat is a modular conversational artificial intelligence system that allows the user to verbally interact with whatever system GeckoChat is incorporated in.   For instance, in the domain of mobile service robots (MSRs™), GeckoChat can provide voice command communication with a pre-defined ambient user.   Through GeckoChat, the platform has way to relate information it observes about both its world and its internal state.   By combining these system-specific capabilities with the standalone functionality of GeckoChat, GeckoSystems, Inc. provides a robust application to increase interactivity and autonomy in any domain where automated speech recognition and synthesis would be beneficial.

GeckoChat is structured in such a way that language independence is implicit.   We plan to call other language capabilities of GeckoChat, GeckoLingua.

For more information on GeckoChat, please see the GeckoSystems, Inc. GeckoChat Discussion.

2.3 GeckoVision (GeckoOccipita)

The human brain centralizes processing of vision in a posterior section of the brain called the occipital lobe.   Vision is perhaps the ultimate feature of biological sensory design and with the possible exception of language, is the most important method for human sensory input from the outside world, and yet, the blind commonly surprise the sighted with their ability to navigate dynamic environments.

GeckoSystems vision for GeckoVision is as an auxiliary system for object recognition.   We take the approach that traditional computer vision is not a necessary component in robot navigation just as it is not a required component for the blind to successfully care for themselves in their daily routines.   However, because we believe that sensory depravation is the bane of most robot control attempts, we use sensor fusion.   We plan to incorporate computer vision techniques to extend our already sufficient navigation scheme as well as to facilitate in end-effector (e.g. hand) control.

3 Downstream Non-Cognitive Subsystems

While the upstream region is involved in relatively abstract, high-level cognitive processing, the downstream area both executes the results of said cognitive processing while maintaining real-time control over actual platform functionality.   It compares most readily with the autonomous nervous system, which is made up primarily of the cerebellum and the peripheral nervous system, including the spinal cord, associated ganglia, and somatic nerves.

3.1 GeckoCerebellum

The cerebellum, in the human case, is an assembly of specialized neurons in the brain proper that controls real-time movement and balance, thus freeing the higher (frontal) areas of the brain for cognitive processing.   GeckoSystems takes a similar approach in solving the real-time control issue.   Because our architecture, like most biological control architectures, is not flat, we allow for certain tight couplings, as needed, between system and subsystem inputs and outputs such that it is transparent to upstream control nodes like those found within GeckoNav.   In concentrating hard real-time control in the downstream segment of the GeckoSystems architecture, the upstream region is free to manage semi-real-time and deliberative tasks requiring cognitive processing, such as mapping and speech.

The GeckoCerebellum has domain over the GeckoGanglia, which are described below.

3.2 GeckoMedulla

The medulla oblongata is located in the lowest portion of the brain proper and serves as both a relay between the cerebrum, cerebellum and spinal cord and as a control center for the autonomous nervous system.   In terms of the GeckoSystems architecture, we use a similar approach in communications between GeckoNav and the GeckoCerebellum.   While not exactly aligned with the human biological architecture, we feel that the GeckoMedulla metaphor captures the essence of the medulla oblongata's functionality.

The GeckoSystems architecture includes a pipeline link between GeckoNav and the GeckoCerebellum, with module communications filters built in each to facilitate use of arbitrary protocols.   Because of the flexibility of GeckoMedulla communications between the upstream and downstream components of the GeckoSystems architecture, the upstream component (i.e., GeckoNav, GeckoChat and GeckoVision) can be housed off-platform via many current and future wireless communications protocols.

3.3 GeckoGanglia

Even in so-called lower animals, ganglia serve as auxiliary "brains" or units that control processes requiring fast reactive response.   For example, the dinosaurs were proported to have several large ganglia along the lengths of their bodies to facilitate quick, reflexive responses.   Some organisms have only ganglia with no central brain.   This distributed processing network still allows these organisms to remain competitive in their respective environmental domains.   In humans, we have some vestiges of ganglia, especially concentrated in or near the spinal column, which may facilitate in sensory and signal processing, especially in the realm of touch.

GeckoSystems relies heavily on this ganglia model for our real-time control systems.   By utilizing programmable micro-controller units and sophisticated electronics at the level of motor control and sensory pre-processing, raw data is converted to a form amenable to higher-level units such as the GeckoCerebellum.   The compression effect of raw data being converted into increasingly symbolic forms is common in sequential processing systems and is reminiscent of video and audio compression techniques; however, we have gone a step further in adapting this paradigm for parallel-distributed processing in robotics.   One of the expressions of this approach is found in our sensor fusion method.   Currently, we have several ganglia between the GeckoCerebellum and both the sensory and locomotion hardware.

4 Advantages of the GeckoSystems Architecture

One of the several advantages of this architecture is the fact that any given component communicates only with adjacent subsystems, not with the entire system.   Thus, for example, GeckoNav need not "know" how to communicate with any of the several GeckoGanglia, it only need communicate with the GeckoCerebellum, which then handles directing the GeckoGanglia (completely transparent to GeckoNav).   In other terms, GeckoNav runs against abstract inputs and outputs definitive actions without concern as to the execution of said actions.   Indeed, by this one illustration, we show the type of simplification in communication between components that enables the GeckoSystems architecture to be intrinsically multi-platform enabled.

Another advantage is in redundancy and error-recovery.   Consider a system based on a flat architecture: if there is a failure in a component due to hardware problems, communications failure, power spikes, etc., control of the platform is lost, which is inappropriate in any mission critical situation.   While these issues, to the degree possible, may be handled in complex error-handling schemes, the overhead cost inherently increases frequently to the point of not being cost-effective.   However, in a distributed architecture such as GeckoSystems', redundancy is implicit.   If, for example, the GeckoCerebellum loses contact with GeckoNav for a given time interval, the GeckoCerebellum will assume emergency control of the robot until such a time as communication with GeckoNav is re-established.   Similar fail-safes are included within each node so that functionality of the entire system is not totally lost upon malfunction of one of the comprising units.   This ensures a higher level of safety for those interacting in the MSR's environment.

Because of the flexibility of the GeckoMedulla in facilitating communication between the upstream and downstream components of the architecture, a complete on-board solution with both components localized is as straightforward as the alternate solution of onboard downstream and offboard upstream components.   Where longevity of battery life is an issue, the upstream components can be housed on an external computer or even a network of computers.   When a self-contained solution is desirable (e.g., in environments with wireless interference or in security applications), both components can exist on microcontrollers and miniaturized microcomputer boards incorporated into the platform.

5 Potential Disadvantages and Rejoinders

In terms of hardware requirements, doesn't a hierarchical system cost more to manufacture than a flat system?

In any distributed architecture, bounding complexity is a problem.   In terms of manufacturing, having several processing and control units raises assembly intricacy.   In GeckoSystems' experience, we assert that a multi-layered MCU/CPU architecture is far less expensive both in power and dollars than a flat CPU architecture.   Early in our research, we considered a flat architecture and deemed it inappropriate for robust and upgradable MSRs.

What about the power requirements for this architecture: doesn't a distributed system require more power and hence adversely affect battery life?

Not always, since the single greatest factor affecting battery life, excluding screens and hard drives, in laptops and portables is clock speed of the CPU.   All of our demonstration videos are of platforms of maximum clock speeds less than 500 MHz.   Our battery life is typically 12-24 hours between charges.   This in part is because our clock speeds are typically one to two orders of magnitude less than what is found in flat architectures.

6 Summary of the GeckoSystems Architecture

The crux of the GeckoSystems architecture for robot intelligence and control is in its hierarchical, distributed design.   In the upstream or sensory chain of processing, mundane (but potentially intensive) processing is done by real-time control units called GeckoGanglia, with the subsequent converted data handed to the GeckoCerebellum, which then performs further processing, makes certain real-time reactive decisions, and sends symbolized data upstream to GeckoNav.   GeckoNav then deliberates and acts based on current data from lower-level nodes, temporal knowledge from past processing, and a priori knowledge.   The downstream or action chain of processing is inititated within GeckoNav, which then sends symbolized command data to the GeckoCerebellum, which then directs lower-level components (ganglia) to act.   We believe that the implementation of this architecture will usher in a new age of practical, cost effective, autonomous and intelligent robots for personal, business and government use.