Animals have evolved to occupy every environmental niche where we might want an autonomous robot to operate. As a result, they provide proven solutions to the problems of navigation, searching and sensing, often in the most difficult of environments. Much desirable robotic behavior has precise analogies in the behavior of underwater animals. The set of behavioral acts that a lobster or lamprey utilizes in searching for and indentifying prey is exactly what an autonomous underwater robot needs to perform to find mines.

In crustaceans, many action patterns such as those that underlie locomotion and feeding can be evoked by stimulation of single neurons or sets of neurons, peptides and/or amines (Bowerman and Larimer, 1974; Harris-Warrick, et al., 1996). Thus,there is a correspondence between units of behavior, their modulation and underlying neuronal components. We have developed a biomimetic control architecture based on these neuronal components and modulatory principles that utilizes object-oriented programming techniques (Ayers et al., 1994; Ayers, 1995). The architecture is based on state sequences that are derived from the analysis of kinematics of behaving animals (Schlichting and Ayers, 1995). Recent advances in smart materials such as shape memory alloys (Deurling, 1994) and MEMS technology make it feasible to construct robots that combine the behavioral control systems described above with the sensors and actuators analogous to those used utilized by organisms that behave in the littoral zone as well as in high flow environments such as rivers and streams.

The locomotory and taxic behaviors of animals are controlled by mechanisms that are conserved throughout the animal kingdom (Kennedy and Davis, 1977; Stein, 1978). The neuronal mechanisms underlieing locomotion were initially established by study of simple animals including lobsters and crabs, insects, sea slugs and worms (Hoyle, 1976; Kennedy and Davis, 1977). These mechanisms have been formalized into a general model, the command neuron, coordinating neuron, central pattern generator model (CCCPG model, Fig. 2; see also Kennedy and Davis, 1977; Stein, 1978). The model is composed of five major classes of components including central pattern generators (Pinsker and Ayers, 1983), command systems (Kupferman and Weiss, 1978), coordinating systems (Stein, 1976), proprioceptive and exteroceptive sensors (Wiersma, 1977) and phase and amplitude modulating sensory feedback (Stein, 1978).

The fundamental governing concept of the CCCPG model is that the motor output that underlies behavior is generated by genotypically specified central pattern generators that are modulated by peripheral exteroceptive and proprioceptive feedback during behavior (Delcomyn, 1982). In other words the central nervous system can generate central motor programs in the absence of sensory feedback. This central pattern generation model differs fundamentally from reflex-chain models where sensory feedback is necessary to specify transitions between different phases of a cyclic behavior (Sherrington, 1906). The central component consists of :

The peripheral component consists of exteroceptive and proprioceptive sensors that provide feedback to the central pattern generators to generate:

The neuronal control mechanisms of insects and decapod crustacea (cockroaches, locusts, lobsters, crayfish and crabs) have been subjected to considerable reverse engineering over the past 25 years, adequate to permit robust synthetic models of their underlieing organization (Beer et al., 1992). In several cases the actual synaptic networks have been established by electrophysiological stimulation and recording (Pearson, 1976; Chirachri and Clarac, 1989). In fact the resulting neuronal circuit based models can achieve much of the complexity that underlies higher order behavior (beacon tracking, Beer, 1991; adaptive walking in different directions, Ayers and Crisman, 1992). These biological models can be readily adapted to robotic control (Brooks, 1992; Beer et. al; 1992; Ayers and Crisman, 1992). We submit that biologically-based reverse engineering is the most effective procedure both to design autonomous underwater robots as well as to establish detailed higher order control schemes for procedures such as remote sensing and mine countermeasures.

A.	B.
C.	D.

Figure 3 Upper Panels: Lobster Sensors. Upper left: Hair fan water current receptor. Upper right: Statocyst balance organ. Lower panels: MEMS Sensors. Lower left: current receptor, Lower right: Inclinometer

Most exteroceptive sensors in crustaceans are derived from modified hair cells. We are fabricating hair like sensors using the NU MEMS process. Fig. 3 shows diagrams of sensors as fabricated using the NUMEM process.Ê The first device consists of a complex cantilever shape.Ê A central beam is ultimately used as the hair.Ê The surrounding two beams are switches of the type already fabricated at Northeastern.Ê On the central beam, the beam is defined with two mechanically weak points as indicated by the indentations.Ê Our approach is to grab the end of the beam and raise it to a new position allowing it to bend and plastically deform at the weak points.Ê In this way, the hair is raised and becomes normal to the surface (Figure 3c).Ê The great advantage of this concept is its simplicity.

With the hair normal to the surface, forces applied to the hair will cause it to deflect as shown in Figure 3c.Ê Because the outer cantilever beams are mechanically connected to the central beam or hair, the outer beams are bent toward the substrate.Ê The contacts at the ends of these beams eventually (with high enough force) contact the counter electrode on the substrate creating a short circuit between them.Ê Therefore, the sensor provides an on-off signal in this configuration.Ê The flow threshold that is detected is proportional to the stiffness of the beams and the length of the switch cantilevers.

Electromechanical devices that code for ranges of pitch and roll by single bits are being developed for both the pitch and roll axis. We use simple microcantelever systems as indicated in Fig.3d. Since these devices are basically switches they will constitute a labeled line code that will interface directly to our controller. The tilt sensor is required for stability control of the lobster. A tilt sensor can be fabricated by attaching a weight to the end of the hair.Ê The tilt sensor can be made to sense various angles by changing the original angle to which the beam is bent (Figure 3d).Ê Further, by orienting several beams at different angles on the substrate, tilt measurement through any solid angle is possible.

One of the fundamental hindrances to the development of biologically-based robots has been the need for an actuator that approximates muscle. Crustacean muscle has mechanical properties that are quite different than typical robotic actuators and these properties need to be modeled in the interface between the controlling "neuronal" signals and realization of joint angular movements. In particular, crustacean muscle is typically modeled as a tension generating component (the contractile fibers) in series with an elastic component (tendons), both of which are in parallel with an elastic component (muscle cell membranes). Neuromuscular control is mediated by motor neurons which differ in the number of muscle fibers that they recruit. The largest motor neurons activate the largest number of muscle fibers and thus generate the largest forces (Davis, 1971). Motor neurons are recruited in the order of size and the contraction force that they cause in the muscle (Fig4a). We have implemented this size principle of neuromuscular recruitment through recruiter objects that directly activate motor synergies.

A.

B. C.

Fig. 4. A. Neuromuscular recruitment network. Circles on the bottom represent elements of the effector synergy that are recruited in the order of size and activate different spatial ranges of the nitinol muscle. B. Nitinol actuator divided into three spatial ranges that correspond the the elements of the effector synergy. C. Antagonistic pairs of actuators controlling a leg joint

Shape memory alloys exhibit properties that can be adapted as a substitute for muscle. Shape memory alloys undergo a state transformation from a deformable state (martensite) to a remembered state (austentite). When formed into wires, one such shape memory alloy (Nitinol) may change its length by as much as 8% during contraction (Deuring et al., 1992). Nitinol has a maximum and minimum range of action specified by its stretched and remembered lengths. Within this range its control can be made proportional, depending upon the portion of wire activated (Fig. 4b). Thus nitinol wires can be activated by the pulse trains generated by our ambulation controller much as crustacean muscle is activated by neuromuscular transduction (Atwood, 1976) and serve as the linear actuator for artificial muscle. These properties map directly on the architecture of the ambulation controller.

A.

B.

C.

We have developed computer controlled video technology for reverse animation and kinematic analysis of animal behavior (Ayers, 1989, Ayers, 1992). This multi-media system allows correlated acquisition of kinematic and electrophysiological data by simultaneously recording behavior in the video signal and electrophysiology on the audio channels of a high resolution digital VCR. We developed extensions to a public domain image analysis program (NIH Image) which include the capability for color based acquisition and image segmentation as well as time-based quantification of kinematic parameters and correlated analog acquisition (ColorImage, Ayers and Fletcher, 1990; Ayers, 1992). This system allows us to measure animal orientation, joint angles from video on a frame by frame basis to establish the detailed movement strategies kinematics of compensatory, orientational and taxic reflexes as well as the underlieing neuromuscular control signals. As a result it has been possible to establish the coordination patterns and control signals underlieing omnidirectional walking (Ayers and Davis, 1977) as well as undulatory swimming (Ayers, 1989).

To directly transit from behavior to robotic controls, we perform finite state analysis task groups that mediate locomotion and searching individually to determine which synergistic sets are operant during different behavioral acts (Schlichting and Ayers, 1996; Ayers, Mehta and Dragich, 1997). Our analysis of the sequencing of these task groups borrows from a technique utilized by astronomers to detect motion of galactic objects. As the analysis proceeds through each frame of the digital movie, the program flashes between temporally adjacent frames of the movie with a brief pause after each cycle. Appendages that are moving the most flash in these projections. A panel of buttons that represent different states of the task groups (eg. elevation vs depression of the chelipeds, etc. are available to the investigator to specify which groups are active By clicking on the appropriate buttons for each frame, it is possible to efficiently quantify the activity of all task groups at high temporal resolution from video tapes of specimens behaving in a variety of situations. These state diagrams are used to establish control sequences for the robots based on the behavior of the model organisms.

Lobster locomotory movements occur principally around three limb joints. We have performed correlated kinematic and electromyographic analysis of the movements of these limb joints and their underlieing control patterns (Fig. 6). Changes in the coordination of these limb movements underlay walking in different directions. Limb muscles can be divided into three basic synergies. An elevation synergy underlies the swing phase of stepping. A depression synergy underlies the stance phase of stepping and counteracts the force of gravity. Bifunctional synergies are either coupled with elevation to generate the swing phase, with depression to generate the stance phase and provide translational propulsive force or are coactivated to hold the joint rigid when it has a postural function (Ayers and Davis, 1977; Ayers and Clarac, 1978).

A.

B.

At the single limb control level, our ambulation controller relies on three major classes of components that control the elevator, depressor, protractor, retractor, extensor and flexor synergies (Figure 7b). The oscillator component is a software clock that regulates the period of stepping as well as the duration of the swing or elevator phase fraction of the stepping cycle. The clock maintains step timing registers which contains the parameter associated with the desired stepping period. At the termination of each step cycle the oscillator loads the step timing registers with the clock tick values associated with the expected end of the elevator phase and the end of the step cycle based on the desired stepping period. During ongoing operation, the oscillator continually compares the processor clock ticKs with the two registers and when the target times are achieved issues state transition messages to the pattern generator, the recruiter, and the neuronal oscillators of any ipsilateral or contralateral governed limbs.

The third major component of the ambulation controller is the recruiter which determines which of the elements of the propulsive force synergies are active. The recruiter responds to the desired period parameter as well as load sensitive feedback and sets or clears the booleans associated with different elements of the active propulsive pool. In our current implementation, each unit within a synergy is represented as a boolean although it could easily be represented as a scalar to provide more detailed amplitude control.

Figure 7b is a graphical realization of the output of the ambulation controller demonstrating the control signals for forward walking. The output is realized as a strip-chart recording of the timing of activity in the different synergies in the same context as electromyograms or nerve recordings. In propulsive force synergies the width of the trace indicates the degree of recruitment within the pool. Notice that during increases in speed the period of stepping decreases while propulsive force synergies are selectively recruited.

In addition to controlling forward and backward walking, our ambulation controller implements omnidirectional locomotion including lateral walking on the leading and trailing sides (Ayers and Davis, 1977). To complete this orthogonal symmetry, we have included directional and recruiting logic for backward, lateral leading and lateral trailing walking. In all cases, the directional logic eliminates the inappropriate synergies by decoupling propulsive force synergies from the elevation/depression oscillation and adds speed dependent recruitment to the propulsive force synergy. The added connections include those that decouple elevation from flexion and depression from extension during lateral leading walking and those that decouple elevation from extension and depression from flexion during lateral trailing walking.

Fig. 8 Nitinol-based Walking Leg Prototype. The joints of the limb are constructed from plastic riser tubing. Nitinol-based muscles are innervated by control wires.

Fabrication of the limbs involves considerations of both the geometry as well as actuator mechanisms. The actual structure and geometry of the limbs will be quite similar to that of the lobster. Actuator mechanisms, however will require selection of alternatives between several approaches. The actual limb movements will not need to be as precise in timing and amplitude as those of manipulators and the limb itself should be quite mechanically compliant. Lobster legs have several joints that do not actively participate in locomotory movements (they are important in other behaviors such as feeding and grooming) and are held stiff during locomotion. They afford considerable compliance during locomotion as most operate in the same planes as the joints that generate locomotory movements.

Behavioral Controller
The higher order control of the ambulatory robot will be mediated by a dedicated microprocessor that both receives sensor data, controls the command systems to mediate behavioral hierarchies and mediates communication with other command and control systems. The brain processor will integrate vectored interrupt from exteroceptive sensors as well as performance feedback from the limbs. It will implement the behavioral hierarchy of the robot (Davis, 1978) and make decisions about behavioral transitions based on exteroceptive sensor input and control information from communications systems. A primary function of the behavioral controller will be to integrate information from sensors. Sensor input will be both interrupt driven as well as polled. For example input from contact receptors (antennae) and current receptors will generate vectored interrupts while inclinometers will be polled. Acoustic and magnetic sensors for object detection and classification will be operated by dedicated microprocessors that will perform signal processing operations and communicate with the behavioral controller.

The overall behavioral hierarchy of the ambulatory robot will be mediated by the behavioral controller. The behavioral controller is a network that mediates the interactions of the command systems for different behavioral acts. This mediation will include resolving incompatible behaviors evoked by coincident stimuli (eg. forward vs backward walking), orchestrating hybrid behaviors (compensatory reflexes to surge which will involve both postural and locomotory components) and sequencing the commands underlieing multi-component behaviors. The low level commands will consist of both locomotory and postural classes. The locomotory commands specify the direction and frequency of stepping. Superimposed postural commands will mediate compensatory reflexes in the roll and pitch planes as appropriate to ongoing inclinometer and current sensor inputs. Higher level behavioral commands will involve navigational as well as phasic compensatory reflexes such as object avoidance and surge compensatory reflexes.

Biologically-Based Undulatory Robots

The swimming behavior of fishes ranges in organization from anguilliform ( relying on lateral axial undulations, Fig. 9a) to carangiform (relying on a flapping tail and/or fins). Anguilliform locomotion is common in eels and lamprey. As described above, we have developed a multi-media analytical system for reverse kinematic analysis of lamprey swimming (Ayers and Fletcher, 1990; Ayers, 1992). Anguilliform swimming results from propagation of flexion waves from the anterior region of the body to more caudal regions (Fig. 9). During anguilliform locomotion, the propagation time of the waves from nose to tail is equal to the period of the undulations so that the body axis typically exhibits an S shape.

Propagating flexion waves alternate on the two sides to generate undulations. The amplitude and timing of the axial undulations are controlled independently (Ayers, 1989). Swimming behavior is controlled on a flexion wave by flexion wave basis. Turning and other maneuvering actions are mediated by modulation of the amplitude of individual flexion waves.

The thrust generated during anguilliform swimming is pulsatile (Fig. 9c). Peaks of thrust are generated as flexion waves propagate to ~65% of body length and are correlated with maximal unbending of the body axis. Our working hypothesis is that the magnitude of thrust is regulated by modulation of axial stiffness by coactivation of musculature on the two sides of the body. Control of axial stiffness is necessary to adapt the speed of locomotion from low search speeds to more rapid pursuit behavior

Numerous physiological studies have demonstrated that undulatory locomotion is generated by segmental central pattern generators (Grillner and Wallen, 1984) that are coordinated by contralateral and ipsilateral coordinating systems (Fig. 2b). The central pattern generators in turn activate motor neurons that are recruited in order of size to grade the intensity of contractions and the resultant lateral flexions (Grillner and Kashin, 1976).

Neural Circuit-Based Controller

The necessary control signals are easily generated with a minor modification of the finite-state machine used for ambulatory control (Ayers and Crisman, 1992). In undulatory systems lateral flexion signals are sent sequentially to segmental muscles, thus the undulatory controller requires only the clock and recruiter levels of control present in the ambulatory controller (Fig. 10a). We are evaluating SMA wires and similar technologies as linear actuators to mediate axial undulations. Nitinol wires of diameters as small as 50µ can generate tensions of up to 30 grams. Arrays of such wires are activated by the controller to generate undulatory movement by sequentially activating flexions over different body regions (Fig. 10). The controller is implemented on a sequential processor which sets and clear digital output signals. The actual control signals will be strobed to a set of shift registers, each bit of which will gate a set of transistors that control current to the actuators (Fig. 11). Thus the actuator control signals will be regularly reset at the time of state changes in the controlling program.

A.

B.

C. D.

An undulatory actuator mechanism is composed of a polyurethane or polyvinyl chloride that has an array of SMA wires affixed on each side. To approximate fish neuromuscular systems wire arrays must be staggered along the body axis so that the most anterior and posterior wires do not overlap. This will allow the body to flex at any point and for flexion waves to propagate. The rod and SMA array can be encased in a molded rubber form.

The undulation controller generates patterns of activation of elements of the nitinol wire arrays that feature a rostro-caudal phase lag and contralateral alternation (Fig. 10c). These patterns of activation will be achieved with the finite state machine controller. Each element of the Nitinol array will be controlled by a separate oscillator and recruiter component. Regulation of thrust is achieved, in principle, by modulation of the spatial range of contraction mediated by the SMA actuators. Modulation of the spatial range is achieved by providing current at different distances along the wire relative to a grounded end.

The robot will be designed to be slightly negatively buoyant, as is the organism. The system will be constructed so that the center of gravity in the roll plane is ventral and the body axis is slightly positively buoyant toward the nose. An electronic compass will be employed to control heading and provide a sense of direction. Through this mechanism the processor will sample the range of headings for several cycles and then evoke modulatory flexion waves to mediate compensatory turning. The undulatory robot can also take advantage of bilaterally paired sonar sensors to mediate homing and navigation toward sonar beacons.

Fig. 11. Serial line-base interface for sensors and actuators. The controller for the biomimetic robots will operate on a embedded microcontroller. Sensor switches are mapped on the bits of shift registers that are routed to an input serial line. Actuator bits for different spatial ranges of the nitinol actuators are routed to shift registers through an output serial line. Sensors are polled regularly while actuator control signals are written during state changes of the controller.

The control circuitry of the robot has been realized on a small single board computer. The controller as presently implemented features few floating point calculations and both the sensor and effector signals are represented as arrays of bits. For example, an inclinometer is represented by one byte, each bit of which represents angular inclination over a particular range of angles. Similarly the shape memory actuators consist of an array of wires of different gauges, each element of which corresponds to a bit. Thus the control of each joint will require a byte. These input and output bytes can readily be implemented by two series of shift registers which are interfaced to the computer through serial line drivers. As the controller program executes it will both read from and write to these shift registers as each state transition occurs. During quiescence, the program might poll the sensor serial line to look for changes in the input or react to sonar control signal generated by a human operator. We have developed a transponder system for acoustic telemetry which could be readily adapted to messaging two and from shallow water robots

The robotic systems we have described are ideally suited to a comprehensive littoral zone mine countermeasure effort. We are integrating an acoustic lane marking system with an acoustic teleoperation system with these autonomous systems to support random search procedures in the littoral zone. In this scheme, the ambulatory robots will perform search on the bottom while the undulatory systems will operate in the water column to address suspended floating mines (Fig. 12). Both systems will rely on high frequency directional sonar transponders to perform initial localization during random searches. The undulatory system will utilize a down-looking sonar transponder to maintain altitude in the candidate altitudes of suspended mines. The robots will have adequate volume to incorporate additional sensors for detection and identification. For example, in the lobster-based system, the claw control surfaces can house magnetic, chemical, acoustic and electronic sensors which can be placed in extremely close proximity to mine candidates. Any ballasting problems can be compensated for with buoyancy elements.

Fig. 12. Mine countermeasures conducted by biomimetic systems. Sonar sensors in claw surfaces can mediate scanning for mine-like objects. A look-down sonar sensor can maintain locomotion at the candidate altitude range of suspended mines. During search procedures walking robots can cover the bottom while undulatory robots can cover suspended mines.

Existing models of the neurophysiological basis of animal locomotion can be readily adapted to robotic control. The use of muscle-like actuators, biomimetic sensors and conserved neuronal network-based controllers can greatly simplify and unify physical implementations of both ambulatory and undulatory robots. These robots can fulfill a variety of missions including remote sensing, ship tagging, and mine countermeasures.