Control in Robotics and Automation

348 CHAPTER 12 / ROBOTICS WITH PERCEPTION AND ACTION NETS

1 INTRODUCTION

Robotic systems aim at achieving intelligence in behavior and dexterity in motion through a

real-time connection between sensing and action. Achieving such intelligence and dexterity

often requires an integration of distributed sensors and actuators to provide a rich source of

sensory and movement patterns that can be clustered into higher levels of concepts and

actions. The key to successful integration may be a system architecture that supports com-

putational requirements unique to robotics, including uncertainty management and adaptive

error recovery through the interaction among such processes as feature transformation and

abstraction, data and concept fusion [1-14], consistency maintenance among data and

knowledge, and monitoring and replanning.

In spite of the fact that a decade of research and development in robotics has produced

numerous theoretical and experimental results, robots are yet to acquire the level of

intelligence and dexterity required for autonomous task execution in unstructured environ-

ments. Conventional approaches to building robotic systems without underlying computa-

tional principles of integrating sensing, knowledge, and action in real time seem to suffer from

limitations in the task complexity they can handle. If robot intelligtence is measured in terms

of a power-to-weight ratio, where the power is defined by the product of the complexity and

execution speed of tasks and the weight is defined by the product of volume and cost

associated with the required hardware and software, an order of magnitude improvement in

the power-to-weight ratio seems necessary for the new generation of robotics. A robot's

intelligence may be manifested by its extended autonomy. However, the extension should not

simply be the result of aggregating additional functional units, which may cause a reduction

of the power or power-to-weight ratio by increasing space and time complexity. It is

necessary to develop a system architecture that supports extended autonomy without a

decrease in the power or power-to-weight ratio. An architecture that embeds system

knowledge as well as a general problem-solving paradigm in itself may be desirable.

Planetary science sampling robots should possess extended autonomy with the capabilities

of uncertainty management, adaptation to new situations, and fault tolerance. To provide the

robot with extended autonomy requires the integration of a high level of discrete event

planning and low level of continuous time control in a hierarchy of multiresolution time

scales. However, such integration should be done under the limitation of computational

power and the requirement of real-time operation. Conventional architectures for intelligent

robotic systems, such as the subsumption architecture [15] and Nasrem architecture [16],

do not directly address the problem of reducing uncertainties as well as dealing with

unexpected events and system faults. Furthermore, the efficacy and efficiency of integrating

planning and control in multiresolution time scales are yet to be consolidated.

An architecture of intelligent robotic systems, referred to here as a perception-action net

(PAN), is presented for planetary robotic sampling. While connecting sensing and action in

real time. PAN automatically synthesizes goal-oriented behaviors or sequences of actions

toward the set goals under uncertainties, errors, and faults, through task monitoring and

replanning.

In this chapter, we present a method of system uncertainty management based on

representing the overall system sensing capabilities by a perception net and propagating

uncertainties and errors forward and backward through the net for consistency by a

geometric algorithm. A geometric fusion method with a statistical basis can be found in the

literature [17]. However, there are number of problems that statistical methods cannot

handle, such as uncertainty propagation in feature transformation when nonlinearity is

involved and treating system constraints for consistency. The perception net is capable of