This article discusses various aspects of underwater vehicles based on biological intelligence. Recent advances in biomimetics have made it feasible to integrate underwater robots that capture the performance advantages of their animal model. In the biomimetic robots, neurons and synapses are represented as structures, and the equations for each element are updated asynchronously at discrete times in a run-time loop. Search vectors composed of a heading and a distance monitored by a compass and visual odometry can organize the autonomous behavior of the vehicles, and these supervisory commands can be transmitted to the vehicle via sonar by a human operator or random search algorithm. The biological intelligence system that is being researched and developed can mediate supervised reactive autonomy through neuronal integrative processes in networks. The networks of biological intelligence represent a viable alternative to the algorithms of artificial intelligence in the achievement of robotic autonomy.
Recent advances in biomimetics have made it feasible to integrate underwater robots that capture the performance advantages of their animal models [1,2]. Animals that live in underwater environments are typically only slightly negatively buoyant so hydrodynamic forces are much stronger than gravitational forces . Although most have visual systems, their ability to use vision is often hampered by low light and/or turbidity. As a result these animals have developed behavioral sets mediated by a limited sensor suite and differ profoundly from terrestrial or aerial animal models .
We have selected two predatory animals as models for our underwater robots (Figure 1). Lobsters are high-end generalist predators that occupy a broad variety of ecological niches ranging from estuaries to beyond the continental shelf . They are among the most maneuverable of arthropods and are able to walk in any direction and change their direction on a step-by-step basis while maintaining hydrodynamic stability.
The sea lamprey is equally at home in both still and flowing water. It is able to pursue fish and then attach itself with its oral sucker both to feed and to migrate. . Most importantly, due to the presence of re-identifiable neurons, both the lobster and lamprey have become canonical neurobiological models and the basis of much of their behavior is understood at the level of neuronal circuits [5,7,8]
At the level of innate behavior, these animals have evolved a self-organized command neuron, coordinating neuron, central pattern generator (CCCPG) architecture that is conserved among all animal groups [9,10]. This architecture is illustrated in Figure 2 and is based on several categories of network elements . The most fundamental are central pattern generators , which are central networks that generate rhythmic behavior in the absence of timing information from the brain and sensory feedback. The central pattern generators are coordinated by intersegmental coordinating neurons that cause a governed CPG to perturb its timing relative to a governing CPG to establish an intersegmental gait . The highest elements in this hierarchy are command neurons that turn on CPGs through parametric modulation and control their average frequency [14,15]. These central networks are modulated by selfgenerated proprioceptive sensory feedback and external exteroceptive sensory feedback (due to gravity or flow) to adapt to contingencies of the environment.
Our initial explorations of biomimetic robot controllers were based on finite state machines that implemented the rules of the CPGs for lobster walking and lamprey swimming [16,17]. Although these systems generated excellent replicas of the behavior of the model organisms, they did not capture the adaptability of the animals to external perturbation. We explored the use of electronic neurons [18,19], which are analog computers that solve the dynamical equations describing living neuronal activity. Electronic neurons generate highly adaptive motor programs (Figure 6), but are physically too large to fit in the biomimetic robots. Although these systems can be implemented in subthreshold analog VLSI circuits, it is difficult to tune the parameters for the equations that determine dynamics .
Discrete Time Map Neuron and Synapse Model
We have settled on a phenomenological model of neurons and synapses [21,22] based on a two dimensional map (Figure 3). As this map represents the dynamics of the membrane potential trajectory, it profoundly reduces the dimensionality of the model, allowing realtime computation on small processors. The equations (Figure 3A-B) define a fast (x) and slow variable (y) that constrain a two-dimensional map of the membrane voltage in cycle n+1 as a function of the membrane voltage in cycle n (Figure 3C). The function is nonlinear over different ranges of xn corresponding to sub-threshold, threshold, and spiking activity (Figure 3B). The nonlinear function (Figure 3A-B) is solved in a discrete-time loop and generates a new neuron voltage xn+1 when passed the present voltage xn, the previous voltage xn-1, and a synaptic current input. These maps define the voltage trajectory as a function of time and define a broad variety of neuronal types (Figure 4). Two control parameters, alpha (α: Figure 3B) and sigma (σ: Figure 3A), determine whether the neuron is silent, spiking, or bursting and its average frequency (Figure 4). The synaptic current (β) is calculated using the following variables: relaxation rate, synaptic strength, postsynaptic voltage, and reversal potential. The topology of the network is specified in the synapse structures that point to pre- and post-synaptic neurons. This architecture allows us to implement a broad range of neuronal integrative mechanisms.
The primary advantage of this model is that it is based on difference equations rather than differential equations so that large networks can execute in real-time on small processors [22,23]. In fact, the neuron and synapse model will execute in real-time on the Lego Mindstorms™ brick . In the Lego implementation, the neurons and synapses are instantiated as LabView instruments. In the biomimetic robots, neurons and synapses are represented as structures and the equations for each element are updated asynchronously at discrete times in a run-time loop.
Central Pattern Generators
To build electronic nervous systems we start with central pattern generators based on the biological networks [25-27]. The lobster CPG (Figure 5A) consists of a 4-component endogenous pacemaker-based neuronal oscillator that generates a 3-phase oscillation (early swing, late swing, and stance). The elevator synergy functions as the neuronal oscillator that determines the frequency of stepping and alternates with the depressor. Two elements of this neuronal oscillator (swing and stance) activate both bifunctional synergies of the basilar joint (protraction and retraction). Walking commands uncouple the inappropriate synergy to generate the forward and backward walking motor programs (Figure 6).
In the lamprey CPG (Figure 5B), there are three pools of about 50 functionally-similar neurons in each hemisegment. CC (contralateral caudal) neurons inhibit the contralateral CC, LIN (lateral interneuron) and EIN (excitatory interneuron) neuronal pools. Here the oscillation is generated primarily by reciprocal inhibition between the hemisegments of opposite sides of the lamprey body, although the EIN neurons have endogenous pacemaker properties .
To complete the electronic nervous system, coordinating neurons (Figure 2) are used to connect the pacemakers of adjacent segments and are tuned to achieve the appropriate gait. Finally the descending commands are connected to the CPGs to bring them into operation (Figure 2). In the lobster walking system there are four commands for forward, backward, lateral leading, and lateral trailing . These can be activated in combinations to achieve diagonal walking. Separate commands operate on the antigravity depressor synergy to control height, pitch, and roll . In RoboLamprey there are commands for slow and fast forward swimming, backward swimming, and amplitude modulation.
The electronic nervous system activates artificial muscle formed from shape memory alloys to control the behavior of the vehicles. The shape memory alloy Nitinol is used to form heart stents (Figure 7A). When formed into wires it can be used as artificial muscle. Contractions are mediated by heating the wire ohmically, which causes it to convert from a deformed martensite state to a more compact austenite state (Figure 7B). When cooled by the surrounding seawater, it becomes deformable and can be stretched by the action of the antagonist muscle by about 5% over its austenite length to the deformed martensite state.
To achieve excitation/contraction coupling, we use a comparator to generate a square wave from the motor neuron action potential (Figure 7C). The square wave in turn gates a power transistor to apply current from a battery through the Nitinol muscle. The current heats the wire to proportionally convert it from the martensite to the austenite state. Thus we use heat in the same way that living muscle uses Ca++ to mediate excitation/contraction coupling . By varying the frequency of motor neuron action potentials, we can achieve pulse width duty cycle control of both contraction velocity and the amplitude of contractions. In RoboLamprey, the actuators form a chevron between segmental “vertebrae” to mediate lateral undulations (Figure 7D). In RoboLobster, pairs of actuators activate the different joints of the leg .
Environmental feedback is sensed by analog sensors and converted to a labeled line code in sensory neurons . To mediate a sense of direction, we use an analog compass (Figure 8A). Here the actual heading is compared with the desired heading to generate a heading deviation statistic (Figure 8C). The heading deviation magnitude is used to modulate a heading deviation (HD) neuron such that it discharges maximally when the heading deviation is 180̊ and is silent when it is 0̊. The HD neuron excites three sensory interneurons that are connected by lateral inhibitory connections with increasing thresholds so that each of the three responds to different ranges of heading deviation.
These sensory interneurons activate command neurons to mediate fusion of exteroceptive reflexes, where several command neurons are activated in parallel and their effects are combined at the CPGs (Figure 9). Primary command neurons project from the “brain” to each of the segmental CPGs and their synaptic and modulatory effects summate there. Analogous commands exist for decapods such as the lobster , but are complicated by the capability for lateral walking in addition to forward and backward walking. Layered exteroceptive reflexes therefore control heading and mediate adaptations to optical and hydrodynamic flow, collisions, and gravity in parallel.
We plan to control the systems by supervised reactiveautonomy. Consider a human taking a dog on a walk. The human is the supervisor and the dog is reactively autonomous. If the human throws a stick in a lake, the dog autonomously fetches it and returns. This is the natural way to control robots in the field and allows a single operator to control multiple robots.
Nature provides excellent lessons of how to mediate supervised reactive autonomy. The waggle dance of bees provides a canonical example . Here guide bees communicate a heading relative to the sun and a distance to the food source. We have implemented a visual odometer in the honeybee platform and this will operate in an analogous fashion in the walking and swimming robots . Search vectors composed of a heading and a distance monitored by a compass and visual odometry can organize the autonomous behavior of the vehicles and these supervisory commands can be transmitted to the vehicle via sonar by a human operator or random search algorithm.
Communications and Localization
We are integrating a long baseline array for communications and localization of the robots . Basic communications are provided by a piezoelectric acoustic transducer (Figure 10A) . Supervisory commands will be transmitted as message packets of four tones delivered at discrete intervals (i.e., 50, 100, 150, etc.). The four tones will define three intervals that specify a target, method, and data. When a vehicle receives and successfully decodes a message it will send an acknowledgement.
emspFigure 10B illustrates how a long baseline sonar array will be configured in the field. Three tracked robots based on the Abalone (Robolone, Figure 10C) will host a cabled float antenna with GPS and sonar transducers. A central base station will integrate batteries, an inductive charging station and a wireless sonobuoy link to shore. The Robolone will form a tri-radiate array, elevate the GPS to the surface to establish an absolute position, then submerge to the optimal depth for sonar communications and these will be transmitted through cables to the base station (Figure 10B). At the end of search vectors, coded tones transmitted by the vehicles will be detected by the Robolone and relative arrival times will be used to compute the position of the vehicle relative to the docking station (Figure 10D). When the vehicle needs to home to dock, it can use a short baseline sonar array (Figure 10E) to compute the deviation of the sonar target from the hull in azimuth and inclination (Figure 10F) to mediate heading deviation reflexes that will allow the vehicle to home on a sonar beacon on the docking station. A key goal of this system will be to achieve energy autonomy and persistence in underwater habitats.
A fundamental premise of this system is that the intrinsic control of the system is based on neuronal synaptic networks. This biological intelligence can mediate supervised reactive autonomy through neuronal integrative processes in networks . The only algorithms will be in the localization and communication systems. These will communicate goals to the vehicle's electronic nervous system, such as the desired heading and synaptic strengths proportional to the desired distance. Exteroceptive reflex networks will use these goals to govern the achievement of the desired path. The networks of biological intelligence represent a viable alternative to the algorithms of artificial intelligence in the achievement of robotic autonomy.