Intelligent identification and control for autonomous guided vehicles using adaptive fuzzy-based algorithms

Fuzzy logic was first suggested as the mechanism by which humans drive cars. This paper addresses the use of fuzzy logic and algorithms towards the intelligent autonomous motion control of land vehicles. To cope with vehicle complexities, internal parametric changes, and with unpredictable environmental effects, the controllers that are presented, whilst heuristic in nature, are self-organizing or self-learning in that they generate automatically by observation an experiential rule base that models the vehicle, and via an appropriate performance index an optimal control rule base that is robust to large parametric changes. The methodology presented is applicable to any complex process which is too difficult to model or control using conventional methods, or which has relied on the experience of a human operator. An overview of fuzzy logic and static fuzzy logic control (akin to expert systems) is provided, together with illustrative examples. INTRO D U CTION The earliest automatically guided vehicles were developed in Japan ~, the USA 2 and the EEC 3. These were characterized by complex supervisory control systems with partial autonomy through on-board navigation. More recently 4, a wide range of special-purpose and generic autonomous guided vehicles (AGV) research programmes have been initiated (in the UK the MOD MARDI programme 5 (1988) includes tracked cross country and road vehicles; in the USA the DARPA programme6; in the EEC the Esprit II project (No. 2483) Panorama, includes a tracked mining vehicle, wheeled crosscountry vehicle and a laboratory test bed). Of central importance to AGVs are the intelligent tasks of: 1. Multi-sensor data integration or fusion 7'8 to locate the vehicle, to represent or model its internal states and its environment (including obstacles) and to assess the current system situation state vector 9 (see also the EEC Esprit I project (No. 1560) SKIDS). 2. Planning and navigation. 3. Motion control s'1°'~1. The systems architecture of the majority of AGVs is hierarchical 12'13 with usually three levels of abstraction (as with command and control systems9). At the highest, most abstract (and least time-critical, precise or detailed) level, is the planner, which operates on a global mission to determine the connected subgoals or tasks to achieve an assigned objective. At the next level, the navigator utilizes a detailed plan or map to, say, evaluate an obstacle-free local path that optimizes a performance criterion (say, minimum fuel usage or minimum path . . . . ) to produce motion and velocity trajectories. Given these trajectories, the pilot or motion controller must provide, in real time, optimal motion control avoiding obstacles not identified by the planning or navigation stages--requiring direct local feedback of 'visually' acquired proximity sensor data and appropriate very local path adaption. Vehicle motion control from current position and velocity is a two-point boundary-value problem with constrained states; by utilizing a 0952-1976/89/040267-1952.00 © 1989 Pineridge Press Ltd Eng. Appli. of AI, 1989, Vol. 2, December 267 Intelligent identification and control for autonomous guided vehicles: C. J. Harris and C. G. Moore hierarchical problem decomposition, the search space for optimization can be reduced and the resulting subproblem of motion control can then be defined over unconstrained subspaces. The procedure is then establishing obstacle-free subspaces and selecting those which provide, say, minimum time or effort trajectories, but since the vehicle and its environmental database are based upon models or representations that are incomplete, uncertain or fuzzy, then the associated rules or controls are equally fuzzy. Humans use fuzzy-like algorithms when they drive or park a vehicle, search for an object or obstacle, etc., adopting procedures that are sufficiently flexible, robust, imprecise and intelligent that they can be adapted to slightly different situations (the principle of generalization) and have a capability to learn. An approximate solution is acceptable in that there are tolerances in the goals as well as in uncertainties associated with 'world' models. This is a natural top-down approach to AGVs since there is imperfect knowledge of (i) the environment (due to incompleteness and uncertainty of sensor data), (ii) the vehicle dynamics (e.g. variations in payload, velocity, frictional forces, road conditions), etc. and (iii) routes to be followed and some task objectives. Zadeh's principle of incompatibility ~4 and fuzzy logic applies particularly to intelligent guidance and motion control of AGVs in unstructured environments, e.g. human drivers utilize experiential models, with inexact observations linguistically modelled as fuzzy sets (such as 'near', 'close' or 'very close') and implement vehicle control through the maximization of some objective (such as minimum time, or minimum fuel usage) via a set of adaptive fuzzy rules or algorithms. Fuzzy decision rule methods have been proposed for path determination of AGVs that progressively discover the environment 14, for car parking15, ~ 6 and motion control 5'~'17 of AGVs. Classical control is based upon (i) deriving from physical laws or identification algorithms a set of model equations, and (ii) generating a set of feedback control laws that ensure that these models behave as desired. To evaluate an AGV dynamic state equation it is necessary to evaluate the longitudinal and lateral dynamics. The vehicle's longitudinal dynamics can be represented TM by the nonlinear equation: mb=Fl(Io, We, Tp, Rs), (1) where m is vehicle mass, v vehicle velocity and FI (.) is a complex function that includes gearing; inertial terms 19 for rotating elements; engine speed We; throttle position Tp; and resistance or loss terms R~ to represent drag, rolling resistance/braking, etc. Similarly the lateral dynamics can be represented by the two degree of freedom 'planar bicycle' model m@=F2(Uf, fr, v, cy, 0~, e, y), (2) I~O~=F3(Ui, f .v , ci, O~,e,Y), where Fz(.), F3(.) are also nonlinear functions of the yaw angle 0r; steering angle e; side slip angle y; wheel circumferential and normal side forces Uy, cy respectively; and fr the coefficient of friction between wheels and surface. I= is the vehicle moment of inertia about the vertical direction. The control inputs to (1) and (2) are acceleration a (=b); and steering angle e. The composite system (1), (2) is both nonlinear and nonstationary (e.g. parameters m,f); also variables cy, Uy, ~ satisfy other complex dynamical relationships 18. However, these equations can be linearized and for a constant velocity vehicle on a smooth small radius of curvature road conventional Kalman filtering and pole placement control methods have been used 1°. It is, however, clear for an unstructured environment with substantial variations in environmental and vehicle states that the classical 'bottom up' control methods for AGVs are inappropriate, and a 'top down' experiential approach is currently the only practically feasible generic approach. The layout of the paper is: first, an introduction and overview of fuzzy logic and algorithms; next, generation of fuzzy rule bases for system modelling and control, followed by a discussion of fuzzy controller synthesis and, finally, a description of self-organizing or learning fuzzy controllers. F U Z Z Y SETS, LOGIC, RELATIONSHIPS AND ALGORITHMS Fuzzy sets A fuzzy set is a more-general form of a classical set. In a classical set, a collection of objects, or points in a space, or attributes, are said to be elements of that set; a particular element of the universe of discourse is either in that set or outside it. However, in fuzzy-set theory, a fuzzy set A of a universe of discourse X (A c X) may have elements x ~ X which partially belong to the set. The degree to which it belongs to the set A is characterized by the membership function/~A(x) in the interval [0, 1] that represents the orade of membership, with 1 representing full membership, 0.5 (etc.) partial 268 Eng. Appli. of AI, 1989, Vol. 2, December Intelligent identification and control for autonomous guided vehicles. C. J. Harris and C. G. Moore