Biol Cybern (2013) 107:497–512DOI 10.1007/s00422-012-0524-4

ORIGINAL PAPER

A computational model of visually guided locomotion in lamprey

Iman Kamali Sarvestani · Alexander Kozlov ·Nalin Harischandra · Sten Grillner · Örjan Ekeberg

Received: 21 March 2012 / Accepted: 24 September 2012 / Published online: 3 November 2012© Springer-Verlag Berlin Heidelberg 2012

Abstract This study addresses mechanisms for the gener-ation and selection of visual behaviors in anamniotes. Todemonstrate the function of these mechanisms, we have con-structed an experimental platform where a simulated animalswims around in a virtual environment containing visuallydetectable objects. The simulated animal moves as a result ofsimulated mechanical forces between the water and its body.The undulations of the body are generated by contractionof simulated muscles attached to realistic body components.Muscles are driven by simulated motoneurons within net-works of central pattern generators. Reticulospinal neurons,which drive the spinal pattern generators, are in turn drivendirectly and indirectly by visuomotor centers in the brain-stem. The neural networks representing visuomotor centersreceive sensory input from a simplified retina. The model also

Electronic supplementary material The online version of thisarticle (doi:10.1007/s00422-012-0524-4) contains supplementarymaterial, which is available to authorized users.

This article forms part of a special issue of Biological Cyberneticsentitled “Lamprey, Salamander Robots and Central Nervous System”.

I. Kamali Sarvestani (B) · A. Kozlov · N. Harischandra ·Ö. EkebergDepartment of Computational Biology, School of ComputerScience and Communication, KTH Royal Institute of Technology,10044 Stockholm, Swedene-mail: imanks@kth.se

A. Kozlov · S. GrillnerDepartment of Neuroscience, Karolinska Institute, Stockholm,Sweden

S. GrillnerNobel Institute for Neurophysiology, Stockholm, Sweden

I. Kamali Sarvestani · A. Kozlov · N. Harischandra · S. GrillnerStockholm Brain Institute, Stockholm Sweden

includes major components of the basal ganglia, as these arehypothesized to be key components in behavior selection.We have hypothesized that sensorimotor transformation intectum and pretectum transforms the place-coded retinalinformation into rate-coded turning commands in the reticu-lospinal neurons via a recruitment network mimicking thelayered structure of tectal areas. Via engagement of thebasal ganglia, the system proves to be capable of selectingamong several possible responses, even if exposed to con-flicting stimuli. The anatomically based structure of the con-trol system makes it possible to disconnect different neuralcomponents, yielding concrete predictions of how animalswith corresponding lesions would behave. The model con-firms that the neural networks identified in the lamprey arecapable of responding appropriately to simple, multiple, andconflicting stimuli.

Keywords Tectum · Pretectum · Basal ganglia ·Mesencephalic locomotor region · Reticulospinal ·Central pattern generator · Lamprey

1 Introduction

One of the main challenges of the neurosciences is to under-stand the processes involved in sensorimotor transformation.The diversity and complexity of connections between the sen-sory system and motor structures complicates the analysis ofsensorimotor transformation. These connections give rise toseveral pathways running partially in parallel. For example,in anamniotes (fish and amphibians), visual pathways fromthe retina and the tectum, olfactory pathways from olfactoryorgans and the olfactory bulb, and cutaneous pathways fromthe trigeminal nerve and trigeminal relay cells all convergeon the reticulospinal neurons (Dubuc et al. 2008).

123

498 Biol Cybern (2013) 107:497–512

Another major challenge when analyzing sensorimotortransformation in vertebrates is to find the mechanismsthrough which the sensory input triggers the proper behaviorin an individual pathway. Each pathway is composed of sev-eral processing stages, and investigating the role of each stagein the final behavioral outcome is of special interest. Ana-mniotes share the main components of sensorimotor trans-formation mechanisms with other vertebrates. The smallernumber of sensorimotor transformation pathways in anamni-otes makes it easier to analyze their sensorimotor transfor-mation systems, thereby shedding light on the correspondingsystems in all vertebrates.

Out of a large repertoire of behaviors, we focus on mod-eling the pathways generating three basic behaviors shownin different forms by all vertebrates: escape, avoidance, andapproach. These three behaviors may be triggered by a vari-ety of stimuli including visual (Saitoh et al. 2007; Ullén et al.1997; Finkenstädt and Ewert 1983), auditory (Korn and Faber2005), lateral line (Korn and Faber 2005), tactile (Viana DiPrisco et al. 1997, 2000), and olfactory (Selset and Doving1980; Johnson et al. 2009) modalities in anamniotes. Wefocus on the visual origins of these basic behaviors. In theescape behavior, the animal moves away from a stimulus itperceives as aversive. This movement typically starts with aninitial turn in the opposite direction of the aversive stimulus,such as a predator, followed by rapid locomotion. Avoid-ance, however, is usually characterized by a slight steeringto position the body away from the line of collision with astimulus to evade such as an obstacle. In the approach behav-ior, the animal steers toward an appetitive stimulus such as aprey. In anamniotes, lesion and stimulation studies proposethe tectum as a visual approach and prey-catching center andthe pretectum as a visual escape center (Ijspeert and Arbib2000; Petreska 2004). We will use this homology to someextent. Different neuroanatomical and functional aspects ofthis homology are reviewed in the discussion.

Visual stimuli are located in the three-dimensional spaceand are detected by sensory neurons arranged in retinotop-ic order, each tuned to detect the existence of the stimulus ina certain location in the visual field. The retinotopical repre-sentation usually extends to higher visual centers such as thetectum. Motoneurons and muscles, however, are not tuned tothe direction of movement. In motoneurons, the direction isimplicitly coded in the recruitment pattern and firing rate ofthe controlling neurons. The higher the firing rate, the morethe contraction of the muscle and the larger the movement.Bilateral stimulation of reticulospinal neurons or the mes-encephalic locomotor region with graded intensities resultsin faster locomotion and, ultimately, change of gait (Grillneret al. 1997; Jordan 1998; Le et al. 2011). One main functionof a visual sensorimotor transformation center is to properlyconvert theplace-codedsensorystimuli into rate-codedmotorcommands.

All three responses considered here are basically followedor accompanied by locomotion. Thus the driving input to thespinal cord can be assumed to be composed of two com-ponents: a symmetric component responsible for locomo-tion and an asymmetric component responsible for steeringtoward or away from the stimulus. The idea of asymmet-ric activation leading to asymmetric motor command hasbeen suggested both in lamprey studies (Zelenin et al. 2001)and synthetic psychology (Braitenberg 1984). A secondmain function of visual sensorimotor transformation centers,therefore, is to generate both the symmetric and asymmet-ric components of the movement. We aim to model distinctvisuomotor centers responsible for escape, avoidance, andapproach that can transform the retinotopically organizedsensory data into symmetric and asymmetric components ofactivation driving the spinal cord.

Visuomotor centers, the locomotor transmission mech-anisms, and action selection circuitry in anamniotes arestudied extensively. However, the mechanisms by which thevisuomotor centers transform place-coded sensory input intorate-coded motor output and the machinery generating anddelivering symmetric and asymmetric drive components toreticulospinal neurons have not been studied much in ana-mniotes. In this study we first develop hypotheses about thesensorimotor transformation mechanisms and the way theygenerate the symmetric and asymmetric components of loco-motion and then use computational modeling to demonstratehow the postulated machinery may collaborate with othersystems involved in action selection (Kamali Sarvestani et al.2011) and locomotion (Ekeberg 1993; Kozlov et al. 2009).Finally, we use our neuromechanical model to suggest sev-eral lesion–stimulation studies that could confirm or falsifythe hypothetical mechanisms.

In constructing our models, we have used data primarilyfrom the lamprey, whose nervous system has been the subjectof many studies (Grillner et al. 2007, 2008). Wherever possi-ble, we have drawn parallels to other anamniotes, especiallysalamanders, which have been studied comprehensively bothin experimental (Sánchez-Camacho et al. 2001a,b, 2002;Finkenstädt and Ewert 1983; Roth and Grunwald 2000;Cabelguen et al. 2010) and computational (Harischandraet al. 2011; Ijspeert 2008; Ijspeert and Arbib 2000; Petreska2004) studies.

2 Methods

To investigate animal behavior, we have used an animat: amodeled animal moving in a virtual environment. The vir-tual environment includes a model of water interacting withthe body surface. Objects representing aversive, evasive, andappetitive stimuli are also included in the virtual environment(Fig. 1).

123

Biol Cybern (2013) 107:497–512 499

Fig. 1 The animat moves around in a virtual environment as a resultof interactions between its body and the water surrounding it. The red,green, and blue objects represent aversive (e.g., predators), evasive (e.g.,obstacles), and appetitive (e.g., prey) stimuli, respectively. (Color figureonline)

Our animat possesses many features of a real animal thatshape its behavior (Chiel et al. 2009) including the physi-cal properties of water being in contact with the animal, themechanical properties of the animal’s body and muscles, theconnectivity of the animal’s spinal cord pattern generators,its descending neuronal control system in visually guidedmovements, and, finally, its action selection mechanisms.Processes involved in object recognition and spatial posi-tioning of stimuli are only modeled on an abstract level.

2.1 Networks

Instantaneous positions of the stimuli in a moving polar coor-dinate system attached to the animat’s head are recorded by its“retina.” The information about aversive, evasive, and appeti-tive stimuli flow in three segregated channels from the retinato visual sensorimotor transformation centers on the oppo-site site (Fig. 2). Direct pathways unilaterally connecting thevisual sensorimotor transformation centers to the reticulo-spinal neurons are responsible for generating the asymmetricdrive in the reticulospinal neurons. By crossing the midline,the approach center activates the reticulospinal neurons of theopposite side. In contrast, escape and avoidance centers pro-ject to the reticulospinal neurons on the same side with strongand weak synaptic strengths, respectively. These synapsesprovide input to reticulospinal neurons, generating weakerresponses in avoidance behavior compared to escape behav-ior. Secondary pathways connect the three visual sensori-motor centers to the reticulospinal neurons via a relay in alocomotor center. The bilateral efferent connections of thelocomotor center in this pathway are in charge of generat-ing the symmetric drive in the reticulospinal population. Thereticulospinal drive activates the spinal locomotor machin-ery, and the animat moves as a result of its interaction with thesurrounding water. The simulation continuously updates the

visual input as the animat changes its position. Figure 2 showsthe gross connectivity of the animat in its virtual environment.

The retina is modeled as three one-dimensional arraysof spiking neurons partitioning the 360◦ of visual field into128 equal size receptive fields (64 on each side). Each arrayrecords the information about direction and distance of onestimulus type so that information about aversive (red), eva-sive (green), and appetitive (blue) stimuli flow in segregatedchannels toward the visuomotor centers. Thus the position(index) of a neuron within a certain array represents the direc-tion of the stimulus, and the firing rate of that neuron repre-sents the salience of the stimulus.

Each of the animat’s three visual behavior centers is orga-nized in three layers: an input layer, a response layer, andan auxiliary layer. The input layers are composed of arraysarranged in a topographical manner so that each neuron inthe input layer receives spikes from a corresponding neuronin the retina.

The response and auxiliary layers are also composed oftopographically organized arrays of neurons. The arrays areequal in size to those in the retina and input layer. The con-nectivity between the input and response layers is differentin the approach center from those in the escape and avoid-ance centers. In the approach center, each neuron i in theinput layer is connected in register to a response neuron withthe same index i in the response layer. In contrast, each neu-ron i in the input layer of the escape and avoidance centers isconnected to a neuron 64 − i in the response layer, yieldinga left-right mirrored representation of the stimulus position.This difference is the origin of different responses that theanimat displays in approach versus escape behavior. Sincethe activities of the response layers depend on the proximityof the stimuli generating them, they are good candidates forgenerating the symmetric drive in the reticulospinal neurons.Therefore, the three response layers converge on the locomo-tor center, a neuronal population responsible for generatingsymmetric locomotor drive. The locomotor center in turn pro-jects bilaterally to the reticulospinal neurons (Figs. 2 and 4).

The neurons in the response layers of each visuomotorcenter inhibit other neurons in the response layer of the samevisuomotor center. Therefore, if there are several stimuli ofthe same nature (e.g., several appetitive stimuli or severalaversive stimuli or several obstacles) present in the virtualenvironment, only the strongest response will remain in theresponse layer of the corresponding visuomotor center. Thismechanism resolves the conflicts within a certain visuomo-tor center. The conflicts between different visuomotor centersare resolved by the arbitration system (see the next section).

As stated earlier, in visually guided movements the place-coded information from the retina must be transformed intorate-coded information in the motoneurons. Several mecha-nisms have been proposed to convert place coding of sensorystimuli into rate coding of motoneurons (Groh 2001). These

123

500 Biol Cybern (2013) 107:497–512

Fig. 2 Top (dorsal) view of neural connectivity in animat in virtualenvironment (internal connections of CPGs not shown). Escape, avoid-ance, and approach centers receive segregated inputs from the contralat-eral retina and project to the reticulospinal neurons, locomotor center,and the arbitration system. The direct projections from the approach cen-ter to the reticulospinal population cross the midline, while those from

avoidance and escape centers remain on the same side. The locomotorcenters project bilaterally to reticulospinal neurons. The reticulospinalneurons in turn project to all segments of the spinal cord. The three visu-omotor centers receive inhibitory projections from the action selectionmechanisms, which suppress some of the conflicting responses theymay generate

mechanisms typically use differential synaptic efficacies ofsynapses formed between the inputs from an array of sensoryneurons converging on a motoneuron. We take a slightly dif-ferent approach (Fig. 3).

The response array uses an auxiliary array of the samesize to convert the place code into rate code in the targetneurons of the reticulospinal population. The auxiliary neu-rons converge on a single reticulospinal neuron as output.Each neuron i in the response array projects to all neuronsin the auxiliary layer with an index j ≥ i (Fig. 3). Twoexamples of this recruitment system are illustrated in Fig. 3.Each layer of the simplified model shown in Fig. 3 contains8 neurons, which is 16 times smaller than the actual sizeused in our animat. In this example, the receptive field ofneuron number 8 is located in front of the animat, and thereceptive field of neuron number 1 is located behind it. Fig-ure 3a demonstrates the effect of an appetitive stimulus infront and slightly to the right of the animat’s view on theneurons in the approach center. Active neuron 7 in the inputlayer stimulates neuron 7 in the response layer, which inturn activates neurons 7 and 8 in the auxiliary layer. Sinceonly two auxiliary neurons are activated, the reticulospinalneuron receives a slight activation and fires with a low rate.This is in fact the proper amount of activity in this casesince the appetitive stimulus is just slightly to the right inthe animat’s view. Figure 3b demonstrates the effect of anaversive stimulus in front of and slightly to the right of theanimat’s view. Because of the mirrored projections betweenthe input and response layers, stimulation of neuron 7 in theinput layer activates neuron 2 in the response layer, which inturn recruits all neurons 2–8 in the auxiliary layer. Therefore,the output neuron in the reticulospinal population receives a

Fig. 3 Structure and intrinsic connectivity in simplified prototypes ofvisuomotor centers. All centers are organized in three layers of equalsize. Each layer is an array of neurons. The neurons in input and responselayers are tuned to specific directions of stimuli. Connectivity betweenthe response and auxiliary layers is shared in all visuomotor centers.Moreover, the response layer neurons of each visuomotor center inhibitother neurons in the response layer of the same visuomotor center. aNeurons in the input layer of the approach center are connected in reg-ister to the corresponding neurons in the response layer. The auxiliarylayer neurons in the approach center are connected to the output reticu-lospinal neurons of the opposite side. b The neurons in the input layer ofthe escape center are connected to the mirroring neurons in the responselayer. The auxiliary layer neurons in the escape center are connected tothe output reticulospinal neurons of the same side

123

Biol Cybern (2013) 107:497–512 501

large activation and fires with a high rate, which is in corre-spondence with appropriate behavioral response in this case,i.e., a sharp turn. Thus, the position of the stimulus in theretinotopic sensory and response layers transforms into thefiring rate of the output reticulospinal neurons. This transfor-mation generates the asymmetric input needed for steeringto the reticulospinal neurons. Although there is little evi-dence for such a connectivity pattern in vertebrates, it isreported that in deep layers of the superior colliculus (homo-log of the tectum in higher vertebrates) a large population ofcoarsely tuned collicular neurons fire before saccades, andthe ultimate amplitude of the saccades depends on the sum-mation of each individual neuron activity within the popula-tion (Lee et al. 1988). Figure 4 shows the side (lateral) viewof the connectivity between different supraspinal neuronalnetworks in the animat. As described earlier, neurons in theresponse layer project to the locomotor center, whereas theauxiliary layer neurons directly project to the reticulospinalneurons.

The spinal locomotor central pattern generators (CPGs)are modeled as pairs of linear networks distributed alongthe spinal cord on its left and right sides. The two networksare coupled via long-range ipsilateral excitatory projectionsand contralateral inhibitory projections. Network connectiv-ity follows the general structure of a previous model (Kozlovet al. 2009), while the network size and cell complexity arereduced to allow for faster simulations. In total, 1,600 inte-grate-and-fire neurons with spike frequency adaptation areused in the animat’s spinal cord. Ten pairs of special non-spiking units integrate local interneuron activity and providefiltered motor output. A swimming locomotor pattern is pro-duced as a wave of left–right alternating oscillations travelingfrom the rostral to the caudal end. The musculomechanicalmodel of the lamprey is built according to previous simu-lations (Ekeberg 1993) with parameter values adapted fromour salamander model (Harischandra et al. 2010).

2.2 Basal ganglia

The three basic visual behaviors are mutually exclusive inthe sense that they cannot be executed simultaneously. Con-flicts between the responses from visual centers are to beresolved by an arbitration system (Kamali Sarvestani et al.2011). The arbitration system in the animat has an input stagerepresenting the subthalamic nucleus (STN), an output stagerepresenting the internal part of the globus pallidus (GPi),and an intermediate stage representing the external part ofthe globus pallidus (GPe).

In our animat there are three neuronal populations in theSTN. These neurons do not have retinotopic organizations.Instead, all neurons in the response layer of each visuomo-tor center converge on a corresponding neuronal population

in the STN. The output stage of the arbitration system is inturn composed of three neuronal populations, each inhibitingall neurons in the response layers of one of the visuomotorcenters.

The intrinsic connectivity between the STN, GPe, and GPiis shown in Fig. 5. The STN neuron representing each of thethree behaviors excites the GPe neuron representing the samebehavior, which in turn inhibits STN neurons representing theother two behaviors. In this way, the strongest STN activ-ity manages to inhibit the other (competing) behaviors viaGPe. A functionally similar but structurally different networkbetween STN and GPi also inhibits the weaker activities inthe STN in favor of the strongest. Here, each STN neuronrepresenting a certain behavior activates GPi neurons repre-senting the other two behaviors. GPi neurons representinga certain behavior in turn inhibit the STN neuron represent-ing the same behavior (Fig. 5). This connectivity pattern hasthe same functional role as the STN–GPe, but the pattern ofneuronal activity in GPi will be the complementary patternof the activity in GPe. This complementary nature is furtheremphasized by the inhibitory connections from GPe to GPi:each GPe neuron representing a certain behavior inhibits theGPi neuron representing the same behavior.

The highest level of control in our model, i.e., the exten-sion system, has direct access to sensory information fromthe retina as well as the output of the competition in thearbitration system (Fig. 5). The input stage of the extensionsystem, i.e., the STR, receives information about the behav-ior chosen by the arbitration system via projections fromthe subthalamic nucleus (Fig. 5). This configuration enablesthe system to learn, promoting or forbidding certain actionsunder certain sensory and contextual conditions. Such a con-text-dependent action selection makes it different from thearbitration system, which selects actions based exclusivelyon their instantaneous intensity. Thus, the extension systemmay in certain situations promote an action that is not trig-gered directly by visually guided movement centers, or itmay suppress an action that is winning the arbitration com-petition otherwise. Such a higher-order control is contextdependent and learned via the experiencing of positive andnegative outcomes of actions taken in the past. The exten-sion system is a Boolean machine composed of striatal neu-rons whose electrophysiological properties makes them anappropriate choice as being conjunction (AND) neurons andpallidal neurons in GPe and GPi whose electrophysiologi-cal properties suggests that they are disjunction (OR) neu-rons. Although there is no evidence of lampreys learningBoolean logic, given the fact that amphibians can learn suchlogic, together with the full machinery of the basal gangliaobserved in lamprey, we believe it is reasonable to suggestthat lampreys can also learn simple logical rules (see resultsof experiments). This prediction, of course, needs furtherexperiments before it can be justified or falsified.

123

502 Biol Cybern (2013) 107:497–512

Fig. 4 The direct connections between the visuomotor centers and thereticulospinal neurons in the animat arise from the auxiliary layers. Theindirect connections between the visuomotor centers and the reticulo-spinal neurons via the locomotor center arise from the response layers

of the visuomotor centers. The response layers also provide input andreceive inhibitory output to and from the arbitration system. The ani-mat’s extension system receives input from the arbitration system aswell as direct visual input from the retina

Fig. 5 Basal ganglia connectivity in animat. The connectivity betweenSTN, GPe, and GPi forms a winner-take-all network whose end result isthe activity of a single neuron representing the strongest instantaneousresponse in STN and GPe and activity of neurons representing the othertwo responses in GPi. The contextual colors (cyan, magenta, yellow,black, and white) combine to influence the winner of the competitionin the STN/GPe/GPi network. Cyan and magenta cues on one side andblack and white cues on the other are linked together by conjunction(AND) neurons of the striatum. Each conjunction in turn is linked to theother via a disjunction (OR) neuron in GPe. The yellow cue negates eachconjunction by local inhibitory connections between striatal neurons.(Color figure online)

2.3 Numerical simulations

We have modeled a lamprey with realistic physical andmechanicalpropertiesofmusclesandbody thatmovesaroundin an infinitely large model aquarium. The forces from thewater surrounding the body are assumed to be proportional tothe square of transversal velocity through the water.

The muscles are approximated by linear actuators wherethe force depends linearly on muscle length, stretch velocity,and neural input intensity. Simple leaky integrate-and-firemodel neurons are used for all neurons in the simulationexcept for neurons in spinal CPGs where spike frequencyadaptation is employed. Synapses in the supraspinal regionsare modeled as elements with constant weights and delaysduring the course of simulation; postsynaptic currents aremodeled as decaying exponentials. Synaptic weights aretuned to produce spiking frequency in the recipient neuronsin a range of 0–100 Hz. Spinal cord synapses have spikefrequency adaptation, as discussed in previous models(Kozlov et al. 2009).

The whole simulation is made in Python using plug-inlibraries for graphical output as well as neural and mechan-ical simulations. Neuronal networks are simulated using thePython interface of NEST (Gewaltig and Diesmann 2007)on a single core. Selected parameters used for model neurons(membrane capacitance, membrane time constant, refractoryperiod, firing threshold, resting potential, resetting poten-tial, and synaptic current rise time constant) are randomlyassigned to avoid artifacts that may result from using identi-cal model neurons. The distribution of parameters is chosento be uniform with a mean equal to the default parameters ofthe model neurons in NEST. The neural networks are simu-lated for 500 ms prior to movement to avoid transients fromnetwork initiation. The supraspinal input to the spinal cordis updated every 200 ms, implying that the animat selectsactions five times per second.

Mechanical simulations are done using the Open Dynam-ics Engine (Russell Smith, www.ode.org), an open-source,high-performance library for simulating rigid body dynam-ics. Simultaneous simulations of both neural and mechan-ical components are thus performed via PyODE, a Pythoninterface to the ODE library, and PyNEST, a Python inter-

123

Biol Cybern (2013) 107:497–512 503

Fig. 6 The animat reactsproperly to individual stimuli bya escaping an aversive stimulus,b avoiding an obstacle, and capproaching an appetitivestimulus

face to NEST. The complete Python codes for simulationscan be found in the supplementary materials. Interestedreaders can set different scenarios by changing the num-ber and location of different stimuli or simulate lesion ofany neuronal population and observe the end results onthree-dimensional videos. Raster plots showing neural activ-ity in all neuronal populations can be produced so thatthe activities related to certain behaviors can be directlyobserved.

3 Results

3.1 Single stimuli

To verify that the animat is capable of producing the threebasic behaviors of escape, avoidance, and approach whenfacing aversive, evasive, and appetitive stimuli, respectively,scenarios with a single stimulus present in the aquarium werefirst tested. Figure 6 shows the results of the three correspond-ing simulations (see videos Ap.mpg, Ob.mpg, and Av.mpgin supplementary materials).

The animat behaves properly when it encounters singlestimuli: it escapes the aversive stimulus in a sharp turn, avoidsthe obstacles in a slight bending, and approaches the appeti-tive stimulus on a smooth curve. Since simulation parametersare drawn from uniform distributions, the animat takes dif-ferent trajectories in different trials with randomized seeds.A measure of the difference between the trajectories is thetime to reach the target in approach scenarios. Time to reachthe target in Fig. 6c has a mean of 5.350 (s) and a standarddeviation of 1.277 (s).

Figure 7 shows a raster plot of neuronal activity in threelayers of the approach center for the scenario shown inFig. 6c. The position of the appetitive stimulus moves inthe visual field as the animal changes its direction duringapproach behavior. This change is reflected in the activity ofthe three layers of the approach center. The current animat isnot equipped with neural networks that generate optokineticresponses and eye movements. Small sudden shifts in the

position of the stimulus are observed since the animat can-not fix its retina on the object as its head oscillates laterally(see Discussion).

3.2 Multiple stimuli of the same nature

Figure 8 shows the behavior of the simulated animat whenexposed to two stimuli of the same nature. In Fig. 8a, theanimat first escapes the closer aversive stimulus (in the lowerpart of the figure). As the animat approaches the second stim-ulus (in the upper part of the figure), it starts to respond tothis second stimulus instead. The animat in Fig. 8b faces twoappetitive stimuli. It successfully chooses the closer stimulusand suppresses the responses generated by the second stim-ulus. The time used to reach the first target (5.350 ± 1.277 s)is not significantly different from the time taken to reach asingle stimulus in the same position (5.475 ± 1.346 s) in 30randomized trials (see supplementary data).

3.3 Multiple stimuli of different natures

When faced with stimuli of different natures, the animatbehaves more or less in the same way as it did with stim-uli of the same nature as it follows one action at a time, thusmaking fast escapes and precise approaches. Figure 9a showsthe trajectory of the movement (see video ApAv.mpg in sup-plementary materials). Figure 9b shows raster plots of dif-ferent nuclei during this sequence of behavior. As mentionedearlier, the information about aversive and appetitive stimuliflows in segregated pathways. Raster plots in red show theactivity of the neurons in nuclei conveying information aboutaversive stimuli, whereas those in blue show the activity ofneurons conveying information about appetitive stimuli. Thissegregation continues to the level of locomotor center andreticulospinal neurons, where place coding is replaced withrate coding.

Just like the raster plots in Fig. 7, the indices of neuronsin the raster plots of the input layers of visuomotor centersrepresent the relative direction of the animat and the stimuli.

123

504 Biol Cybern (2013) 107:497–512

Fig. 7 Raster plots of neuronal activity in different layers of approachcenter of animat in an approach task. As the animat changes its direc-tion, different neurons report the position of the appetitive stimulus.

This change should have been monotonic for a single stimulus. How-ever, sudden jumps can be observed due to head oscillations in theabsence of compensatory eye movement

The position of the active neurons in the input layers var-ies as the animat moves (Fig. 9b). These variations causemirrored (in escape center) and parallel (in approach cen-ter) activation of the response layer neurons. The number ofactive neurons in the auxiliary layers depends on the positionof the active neurons in the response layers (Fig. 9b). Allauxiliary neurons with a larger index than that of the activeneurons in the response layers are recruited. The differen-tial recruitment of neurons in the auxiliary layers results ingraded activity in the recipient reticulospinal population. Inthe first phase of its movement (time steps 0 to 4 s), the ani-mat exclusively responds to the aversive stimulus. In the sec-ond phase of movement (time steps 5 to 10 s), the escaperesponse vanishes and the stronger approach response dom-inates. Although both stimuli activate the input layers of thecorresponding visuomotor centers throughout the simulation,the response layers are active only in certain phases of move-ment. The dominant response in each phase of movementtotally inhibits the weaker response(s) as a result of activityin the arbitration system (Fig. 9b). In the nuclei of the arbi-tration system, i.e., the STN, GPe, and GPi, neurons withindices 0, 1, and 2 represent approach, avoidance, and escapebehaviors, respectively.

Fig. 8 Internal mechanism of each visuomotor center selects a sin-gle stimulus out of several stimuli of the same nature. a Two aversivestimuli. b Two appetitive stimuli

3.4 Extension system

As opposed to the research on lamprey neuroanatomy, neu-rophysiology, and behavior, studies investigating its learn-ing processes are not abundant. However, the lamprey has a

123

Biol Cybern (2013) 107:497–512 505

Fig. 9 The animat can choose between actions originating in differentvisuomotor centers. a The animat first escapes the aversive stimulus, andonly when it is far enough from the aversive stimulus does it start theapproach behavior toward the appetitive stimulus. b Raster plots show-

ing neuronal activity in different nuclei of the animat’s neural system.The abscissae represent time; the ordinates denote neuron indices. InSTN, GPe, and GPi, indices 0, 1, and 2 represent approach, avoidance,and escape behaviors, respectively

full machinery of the basal ganglia (Stephenson-Jones et al.2011), which may indicate that lampreys should be capableof learning simple motor tasks. Therefore, designing a learn-ing experiment is just speculative. Using the hypothesis thatthe extension system functions as a general Boolean logicmachine (Kamali Sarvestani et al. 2011), we designed anexperiment to test this capability.

In this experiment, the animat has learned that the com-bination of certain color stimuli, either cyan and magenta orblack and white, can neutralize the nature of appetitive stim-uli. Moreover, the presence of another stimulus, the yellowstimulus, can restore the nature of appetitive stimuli even inthe presence of cyan and magenta or black and white stim-uli (Fig. 10 and video CyMa.mpg in supplementary mate-

123

506 Biol Cybern (2013) 107:497–512

Fig. 10 The extension system is capable of associating complex stim-uli to certain behaviors. The animat in this simulation has learned thatif cyan AND magenta stimuli together (c) OR black AND white stim-

uli together (d) are presented, it should avoid the otherwise appetitiveblue object. The animat has, moreover, learned that the yellow stimulusrestores the appetitive nature of the blue object (e)

rials). These stimuli represent contextual cues an animalmay face in its environment (such as day and night or thepresence of a certain odor). Figure 10a shows the behaviorof the animat when the cyan stimulus is presented. How-ever, this activity is not enough to trigger the new behav-ioral mode in the animat. Figure 10b shows the same sce-nario for activated magenta stimulus. This stimulus alonecannot change the behavior of the animat either. The com-bination of these two stimuli (Boolean AND) in Fig. 10c,however, changes the behavior of the animat. The animatneglects the appetitive stimulus and keeps escaping the aver-sive stimulus. Figure 10d shows the same scenario for com-bination of black and white stimuli. Activating either com-bination (Boolean OR) alters the behavior of the animat.Activation of the yellow stimulus in Fig. 10e suppressesthe behavior-changing effect of cyan–magenta and black–white combinations, thereby restoring the normal behaviorof the animat toward the appetitive stimuli. The animat hasthus learned a complex rule: If “cyan AND magenta” OR“black AND white” BUT NOT “yellow” is activated, thenneglect the appetitive stimuli. This is achieved by activeinhibition from the output stage of the extension system(GPi) on the approach center. The restoration of behav-ior is achieved by inhibition of the indirect pathway neu-rons on direct pathway neurons in the input stage of theextension system (for details see Kamali Sarvestani et al.2011).

4 Simulated lesions

4.1 Basal ganglia lesions

By disconnecting the input stage of the extension system (thestriatum) from the rest of the network (thereby mimickinglesioning in a real animal), the animat behaves naturally, but

it is incapable of learning to behave differently under certaincontextual circumstances.

Figure 11 shows the results of disconnecting the arbitra-tion system from the rest of the network. The animat behavesnormally when there is only one type of stimulus in the aquar-ium but faces major problems in behaving properly when twostimuli of different natures are presented together. Since thefiring rate of the reticulospinal neurons is the summation ofall rates suggested by visual guidance centers, the animat,which has lost the ability to suppress weaker responses infavor of the strongest, sums up the turning commands regard-less of their origin. The resultant turn is not toward or awayfrom any one of the stimuli. The animat does not even takean average action (see Discussion).

4.2 Lesions of locomotor center and reticulospinal neurons

The tectoreticular axons of the animat in Fig. 12 have beenlesioned. The animat reacts to all types of stimuli with a for-ward locomotion. The closer the stimulus to the animat, thefaster it will swim forward. The steering command is totallyabolished in this scenario. The animat is still performing loco-motion because the connections from the visuomotor centersto the locomotor center and, ultimately, to reticulospinal neu-rons are intact.

The projection from the locomotor center to the reticu-lospinal neurons have been ablated in the animat shown inFig. 13. Here, the animat is capable of steering but does nothave locomotive support to move forward. Therefore, it steerstoward the proper direction in its place.

4.3 Lesions of tectum

The effect of gradual lesioning of the auxiliary layer of thetectum is investigated in the animat shown in Fig. 14. Grad-ual micro lesioning the neurons in the auxiliary layer of theapproach center results in a gradual decrease in the steering

123

Biol Cybern (2013) 107:497–512 507

Fig. 11 An animat without an arbitration system tends to show exag-gerated responses to several stimuli of different natures. This is basicallybecause the responses add up in the reticulospinal population, therebygenerating strong steering movements

Fig. 12 A lesion to the direct connections between the visuomotorcenters and reticulospinal neurons of the animat removes its steeringcapability but leaves the locomotion untouched since the connectionsfrom the locomotor center are still intact

intensity, requiring the animat to take longer to reach its tar-get. The lesioning is performed in three steps. In the first step,one-third of the neurons in the auxiliary layer are lesioned.In the next two steps, two-thirds and all neurons are lesionedrespectively. The micro lesioning process yields the sameresults in the opposite direction (left to right or right to leftin Fig. 3).

When the same gradual micro lesioning strategy is appliedon the response layer, the results are different: first, thechange in steering intensity abruptly drops from a normalresponse to no response, and second, micro lesioning in the

Fig. 13 A lesion to the connection between the locomotor center andthe reticulospinal neurons keeps the orienting behavior while diminish-ing the locomotion

Fig. 14 Lesioning the auxiliary layer of the approach center in threesteps gradually weakens the steering command. Performing the threesteps of lesion in reverse order yields the same results

opposite direction changes the minimum extent of lesionneeded to remove the response (Fig. 15). As opposed to theauxiliary layer where a population of neurons code the posi-tion of the stimulus, only one neuron in the response layerrecords the place of the appetitive objects. Lesions that donot include this very neuron do not affect behavior.

5 Discussion

We have run simulations of an animat mimicking the behav-ior of a real animal facing different types of stimuli. Theanimat is capable of putting priorities on actions either on

123

508 Biol Cybern (2013) 107:497–512

Fig. 15 a Lesioning the response layer of the approach center in threesteps suddenly diminishes the steering command after one-third lesion-ing. b Performing the three steps of lesioning in reverse order yieldsdifferent results since the steering command is removed only after two-thirds lesioning

the basis of the strength of those actions or on the basis ofthe contextual stimuli suggesting or banning certain behav-iors.

The animat moves by virtue of the undulations of its bodyand the interaction of the body with surrounding water. Theundulations are generated by locomotor pattern generatorslocated along the spinal cord, which are in turn controlled bysupraspinal centers. The combination of visual sensorimotortransformation centers, action selection networks, and spinallocomotor generators with the mechanical model generatingmovements in interaction with the environment provides uswith an effective tool to assess the neural substrates of ani-mal behavior in action. Using this model, one can observeand trace the processes ultimately leading to generation andselection of the appropriate actions. Moreover, by makingsimulated lesions, we have provided some clues about howexperimental lesions or stimulation studies may confirm orfalsify the hypotheses developed in this paper. Since themodel includes both the raster plots of the activity in differentnuclei and the ultimate behavior (supplementary materials),such confirmations or falsifications can be made in detail.

Some level of neurophysiological realism is used in themodel by adopting spiking neuron models in the simula-tions. Although the neuronal model used, i.e., leaky integrate-and-fire neurons (Lapicque 1907), can model input currentswith some good precision, it does not mimic the exact ionicproperties of the membrane. This choice is justified in theabsence of detailed knowledge about single cell properties ofthe supraspinal neurons in anamniotes. More detailed singlecell properties may be added to the model in the future whenbiological data are provided. However, since the parametersof the neuron models used are chosen from random distribu-tions, the model is not critically dependent on any specificneuronal property.

There are several anatomical structures in the real ani-mal resembling the neuronal populations we have used inthe animat. The lamprey pretectum is known to receive inputfrom the contralateral retina and project directly and indi-rectly to reticulospinal neurons (Jones et al. 2009; De Miguelet al. 1990; Kennedy and Rubinson 1977; Kosareva 1980;Vesselkin et al. 1980). The pretectum sends projections tothe ventral thalamus (Nieuwenhuys and Nicholson 1988)and receives inhibitory input from homologous structures ofthe pallidum in both amphibians and lamprey (Maier et al.2010; Stephenson-Jones et al. 2011), which makes it fit thearbitration system we have proposed. The lamprey pretectumhas been suggested as a center mediating aversive behaviors(Ullén et al. 1995, 1997). These similarities suggest it as alikely candidate for the escape center used in the animat.

The lamprey tectum is also known to receive input from thecontralateral retina (Kosareva 1980; Robertson et al. 2006;Jones et al. 2009) and project directly (Zompa and Dubuc1998a,b) and indirectly via mesencephalic locomotor region(MLR; Ménard et al. 2007) to the reticulospinal neurons. Thisstructure is known to be organized in several layers (Heieret al. 1948). The direct connections to the reticulospinal neu-rons originate in the stratum fibrosum profundum (Heier et al.1948), whereas the indirect projections to reticulospinal neu-rons originate in the stratum griseum periventriculare, stra-tum “album” centrale, and stratum griseum central (Ménardet al. 2007), which is similar to the connections used in theanimat. The tectum sends projections to the ventral thala-mus in lamprey and salamander (Roth and Grunwald 2000)and receives inhibitory input from homologous structures ofthe pallidum in lamprey (Stephenson-Jones et al. 2011). Thisconnectivity makes it a proper choice for the suggested arbi-tration system. The tectum is believed to be responsible forapproach and prey-catching behaviors in lamprey (Gahtanet al. 2005) and salamander (Finkenstädt and Ewert 1983).Stimulation of the lamprey tectum results in both eye move-ments and head-orienting behavior (Saitoh et al. 2007). Thesesimilarities suggest it as a likely candidate for the approachcenter used in the animat. However, when addressing thetectum, special care should be taken regarding the specificregion being considered.

Figure 16 shows a schematic view of lamprey mesenceph-alon and the mesencephalic–diencephalic border. Regions Aand B receive input from the anterior retina (Jones et al. 2009)and project bilaterally to reticulospinal neurons (Zompa andDubuc 1998a). Stimulation of region A results in forwardlocomotion (Saitoh et al. 2007). Thus, this region is a can-didate for an “escape forward region” receiving input frommost lateral and posterior regions of the visual field wherepredators probably appear. Since the retinal area projectingto this region is not densely populated, upon observing astimulus in this region, the animal probably swims forwardregardless of the nature of the stimulus to avoid the risk of

123

Biol Cybern (2013) 107:497–512 509

Fig. 16 Dorsal view of lamprey mesencephalon and the mesence-phalic–diencephalic border. Different areas of lamprey tectum receiveinput from different regions of the retina, project differentially to thereticulospinal neurons, and are involved in different behavioral tasks

being predated. Results of stimulating region B suggest thatthe animal moves its head and eyes rhythmically to betterdetect the stimulus.

Region C receives input from the anterior dorsal andanterior retina (Jones et al. 2009) and projects denselyto ipsilateral reticulospinal neurons (Zompa and Dubuc1998a). Stimulation of this region elicits orienting move-ments (Saitoh et al. 2007). The structure of its connectionssuggests that this region is an escape/avoidance region, butsince the retinal region projecting to it is not densely popu-lated, it may not be an acute avoidance center. The posteriorregion of the retina, which possesses the highest density ofganglion cells and covers the visual field in front of the retina,projects to regions D, E, and F (Jones et al. 2009). Region Dprojects mainly contralaterally to the reticulospinal neurons,whereas regions E and F project mainly ipsilaterally to theseneurons (Zompa and Dubuc 1998a). These three regions aretherefore the candidates for being the approach, avoidance,and escape centers modeled in this paper. The amplitude ofthe response generated by stimulating region E is meaning-fully less than the amplitude of the response generated bystimulating the rest of the tectum (Saitoh et al. 2007). There-fore, it can be regarded as the avoidance center in the ani-mat. As mentioned earlier, region F is in the general area ofthe pretectum on the border between the mesencephalon anddiencephalon. Stimulating region G elicits a downward shiftof the eyes and neck, which may make it a candidate as adownward escape region. We have not modeled regions Band G in the current model.

Little is known about the intralaminar structure of the tec-tum and the pretectum in lamprey. There are no anatomicalstudies supporting the layered structure of the pretectum orabout the retinotopy in this structure. However, even the crud-est form of escape triggered by auditory/vibrational stimuligenerated by Mauthner cells is directional (Eaton et al. 2001).

Moreover, studies on larval lamprey show retinotopy in boththe tectum and the pretectum (Cornide-Petronio et al. 2011).Therefore, the existence of retinotopic maps on the pretectumand the rostral tectum and directional sensitivity of escapeand avoidance behaviors in lamprey is very likely.

Both the tectum and the pretectum receive their majorinputs via the contralateral retina. This input is segregatedfrom the retinal level so that different sets of retinal neuronsproject to the tectum and the pretectum, respectively (Joneset al. 2009).

The tectum is believed to be involved in both eye and bodymovements. In this study we have only addressed its role inneck and body movements. This has resulted in unwantedfluctuations in the position of the stimuli on the retina that arefollowed by corresponding changes in the responses. Imple-menting optokinetic responses in the model can improve theperformance of the animat by providing it with fixation andvisual inspection capacities.

The MLR has been suggested as a locomotor center inthe midbrain of almost all classes of vertebrates Grillneret al. 1997; Jordan 1998. The exact extent and componentsof this structure are debated. Stimulation of MLR initiateslocomotion in vertebrates. MLR is known to receive inputfrom the tectum and project to the reticulospinal populationbilaterally (Brocard et al. 2010). These features make it agood correlate for the locomotor center in our animat. It hasbeen suggested that structures in close affiliation with MLRsuch as the pedunculopontine nucleus may contribute to mus-cle tone adjustments in higher vertebrates (Takakusaki et al.2004; Takakusaki 2008). The function we have assumed forMLR in lamprey closely resembles that of muscle toneadjustment.

It has also been suggested that lamprey MLR may be ableto convert short stimulus signals to long duration responses(Smetana et al. 2010). Functionally, this will serve the con-sistency of actions selected even when the stimulus is nolonger visible. This is a very important function to inves-tigate. In this report, however, since we have assumed 360◦vision for the animal, visual stimuli are detectable in any bodyconfiguration, thereby providing a substitute mechanism forconsistency in behavioral response.

It has recently been reported that the lamprey has fullbasal ganglia machinery including striatal substance P andenkephaline-containing neurons, subthalamic nucleus andcells representing pallidal substance P, and enkephaline-receiving neurons (Stephenson-Jones et al. 2011). This find-ing suggests that all vertebrates have complete structures ofthe basal ganglia, possibly with an enrichment of neuronsin terms of their number, diversity, and functionality towardhigher vertebrates.

It has been suggested (McHaffie et al. 2005) that via theirextensive connections with subcortical areas such as the supe-rior colliculus (higher vertebrate homolog of the tectum)

123

510 Biol Cybern (2013) 107:497–512

and pedunculopontine nucleus, the basal ganglia can controlanimal behavior. The novel arbitration-extension hypothesis(Kamali Sarvestani et al. 2011) seems to fit the connectivitypattern and functional organization of the lamprey. The lam-prey tectum and pretectum (among other sensorimotor trans-formation centers) project to the ventral thalamus and receiveinhibitory input from the putative lamprey pallidum, thusqualifying the ventral thalamus and pallidum as the arbitra-tion system. As opposed to the connectivity shown in Fig. 4,the lamprey striatum receives retinal input only via a relay inthe dorsal thalamus. We did not include the dorsal thalamicrelay in our animat since, as stated earlier, we abstracted thevisual processing in favor of motor functions.

Our experiments on the animat with the arbitration systemdisconnected suggest that an animal without action selec-tion can escape imaginary predators or approach nonexistingprey. This is primarily due to the additive effect of severalstimuli on reticulospinal neurons. The “sum of effects” canbe modified to an “average of effects” by adding a popula-tion of normalizing inhibitory (inter)neurons (Groh 2001).Such normalizing inhibitory networks can be found within acertain structure. For example, tectal and pretectal neuronsmay exert inhibitory influences on each other via inhibitoryinterneurons (Meredith and Ramoa 1998; van der Want et al.1992). On the other hand, normalization of activity betweencenters generating different behaviors generated by the samemodality or different modalities needs long-range inhibitorynetworks between different structures. Direct competitionbetween such behavioral centers is probably not an optimalsolution (Redgrave and Prescott 1990). In this sense, the arbi-tration system can be seen as serving as a normalizing systemnormalizing the activity between different structures when aclear winner is not recognizable to the action selection sys-tem.

Anamniotes possess some direct conflict resolution mech-anisms. For example, the pretectum is known to inhibit thetectum with direct projections (Robertson et al. 2006). Thistype of direct inhibition between behavioral centers is, how-ever, more of an exception than a rule. Such an exception mayhave evolved as a result of the very high risk of loss whenfacing a predator in comparison with the chance of gain whenfacing prey.

The reticulospinal neurons are located in a critical posi-tion where they can simultaneously mediate two functionallyimportant tasks: first, mixing the input from different behav-ioral centers and, second, changing the place-coded sensoryinformation to rate-coded information needed for spinal loco-motor activity. Our simulations show that these seeminglysimple features may lead to complicated consequences inanamniote motor control and action selection. Whether or notthis explains why no vertebrate is found without full basalganglia machinery needs further consideration and experi-mentation.

Acknowledgments This work was supported by grants from the EULAMPETRA project, FP7 ICT-2007.8.3., and the Swedish ScienceResearch Council.

References

Braitenberg V (1984). Vehicles: experiments in synthetic psychology.MIT Press, Cambridge

Brocard F, Ryczko D, Fénelon K, Hatem R, Gonzales D, Auclair F,Dubuc R (2010) The transformation of a unilateral locomotorcommand into a symmetrical bilateral activation in the brainstem.J Neurosci 13(30(2):523–533

Cabelguen JM, Ijspeert A, Lamarque S, Ryczko D (2010) Axial dynam-ics during locomotion in vertebrates lesson from the salamander.Prog Brain Res 187:149–162

Cornide-Petronio ME, Barreiro-Iglesias A, Anadón R, RodicioMC (2011) Retinotopy of visual projections to the optic tec-tum and pretectum in larval sea lamprey. Exp Eye Res 92(4):274–281

Chiel HJ, Ting LH, Ekeberg Ö, Hartmann MJ (2009) The brain in itsbody: motor control and sensing in a biomechanical context. JNeurosci 29(41):12807–12814

De Miguel E, Rodicio MC, Anadon R (1990) Organization of thevisual system in larval lampreys: an HRP study. J Comp Neurol302:529–542

Dubuc R, Brocard F, Antri M, Fénelon K, Gariépy JF, Smetana R,Ménard A, Le Ray D, VianaDi Prisco G, Pearlstein E, SirotaMG, Derjean D, St-Pierre M, Zielinski B, Auclair F, VeilleuxD (2008) Initiation of locomotion in lampreys. Brain Res Rev57(1):172–182

Eaton RC, Lee RKK, Foreman MB (2001) The Mauthner cell and otheridentified neurons of the brainstem escape network of fish. ProgNeurobiol 63(4):467–485

Ekeberg Ö (1993) A combined neuronal and mechanical model of fishswimming. Biol Cybern 69:363–374

Ekeberg Ö, Grillner S (1999) Simulations of neuromuscular controlin lamprey swimming. Philos Trans R Soc Lond B Biol Sci354:895–902

Finkenstädt T, Ewert JP (1983) Visual pattern discrimination throughinteractions of neural networks: a combined electrical brain stim-ulation, brain lesion, and extracellular recording study in Salam-andra salamandra. J Comp Physiol 153-1:99–110

Gahtan E, Tanger P, Baier H (2005) Visual prey capture in larvalzebrafish is controlled by identified reticulospinal neurons down-stream of the tectum. J Neurosci 25(40):9294–9303

Gewaltig MO, Diesmann M (2007) NEST (Neural Simulation Tool).Scholarpedia 2(4):1430

Grillner S, Georgopoulos AP, Jordan LM (1997) Selection and initiationof motor behavior. In: Stein PSG, Grillner S, Selverston AI, Stu-art DG, (eds) Neurons, networks, and motor behavior. MIT Press,Cambridge, pp 3–19

Grillner S, Kozlov A, Dario P, Stefanini C, Menciassi A, Lansner A,Hellgren Kotaleski J (2007) Modeling a vertebrate motor system:pattern generation, steering and control of body orientation. ProgBrain Res 165:221–234

Grillner S, Wallén P, Saitoh K, Kozlov A, Robertson B (2008) Neuralbases of goal-directed locomotion in vertebrates – an overview.Brain Res Rev 57(1):2–12

Groh JM (2001) Converting neural signals from place codes to ratecodes. Biol Cybern. 85(3):159–165

Harischandra N, Cabelguen J-M, Ekeberg Ö (2010) A 3D musculo-mechanical model of the salamander for the study of differentgaits and modes of locomotion. Front Neurorobot 4:112

123

Biol Cybern (2013) 107:497–512 511

Harischandra N, Knuesel J, Kozlov A, Bicanski A, Cabelguen JM,Ijspeert A, Ekeberg Ö (2011) Sensory feedback plays a significantrole in generating walking gait and in gait transition in salaman-ders: a simulation study. Front Neurorobot

Heier P (1948) Fundamental principles in the structure of the brain.A Study of the brain of Petromyzon fluvilatis. Acta Anat [Suppl]VI:213

Ijspeert AJ (2008) Central pattern generators for locomotion control inanimals and robots: a review. Neural Netw 21(4):642–653

Ijspeert AJ, Arbib M (2000) Visual tracking in simulated salamanderlocomotion, from animals to animats. In: Meyer JA, Berthoz A,Floreano D, Roitblat H, Wilson SW (eds) Proceedings of the 6thinternational conference on the simulation of adaptive behavior(SAB2000). MIT Press, Cambridge, pp 88–97

Kamali Sarvestani I, Lindahl M, Hellgren-Kotaleski J, Ekeberg Ö(2011) The arbitration-extension hypothesis: a hierarchical inter-pretation of the functional organization of the basal ganglia. FrontSyst Neurosci 5:13

Johnson NS, Yun SS, Thompson HT, Brant CO, Li W (2009) A syn-thesized pheromone induces upstream movement in female sealamprey and summons them into traps. Proc Natl Acad Sci USA106:1021–1026

Jones MR, Grillner S, Robertson B (2009) Selective projection pat-terns from subtypes of retinal ganglion cells to tectum and pretec-tum: distribution and relation to behavior. J Comp Neurol 517(3):257–275

Jordan LM (1998) Initiation of locomotion in mammals. Ann NY AcadSci 860:83–93

Kennedy MC, Rubinson K (1977) Retinal projections in larval, trans-forming and adult sea lamprey Petromyzon marinus. J Comp Neu-rol 171:465–480

Korn H, Faber DS (2005) The Mauthner cell half a century later:a neurobiological model for decision-making. Neuron 47((1):13–28

Kosareva AA (1980) Retinal projections in lamprey (Lampetra fluvia-tilis). J Hirnforsch 21(3):243–256

Kozlov et al. (2009) Simple cellular and network control principlesgovern complex patterns of behavior. PNAS

Lapicque L (1907) Recherches quantitatives sur l’excitation électriquedes nerfs traitée comme une polarisation. J Physiol Pathol Gen9:620–635

Lee C, Rohrer WH, Sparks DL (1988) Population coding of saccad-ic eye movements by neurons in the superior colliculus. Nature332(6162):357–360

Le Ray D, Juvin L, Ryczko D, Dubuc R (2011) Supraspinal controlof locomotion: the mesencephalic locomotor region. In: GossardJP, Dubuc R, Kotla A (eds) Breath, walk and chew, the neuralchallenge: II. Prog Brain Res

Maier S, Walkowiak W, Luksch H, Endepols H (2010) An indirectbasal ganglia pathway in anuran amphibians?. J Chem Neuroanat40(1):21–35

McHaffie JG, Stanford TR, Stein BE, Coizet V, Redgrave P (2005) Sub-cortical loops through the basal ganglia. Trends Neurosci28:401–407

Ménard A, Auclair F, Bourcier-Lucas C, Grillner S, Dubuc R (2007)Descending GABAergic projections to the mesencephalic loco-motor region in the lamprey Petromyzon marinus. J Comp Neurol501(2):260–273

Meredith MA, Ramoa AS (1998) Intrinsic circuitry of the superior col-liculus: pharmacophysiological identification of horizontally ori-ented inhibitory interneurons. J Neurophysiol 79(3):1597–1602

Nieuwenhuys R, Nicholson C (1998) Lampreys, Petromyzontoidea. In:Nieuwenhuys R, ten Donkelaar HJ, Nicholson C (eds) The centralnervous system of vertebrates. Springer, Heidelberg, pp 397–495

Petreska B (2004) A neural visuomotor controller for asimulated Sala-mander Robot. Dissertation. École Polytechnique Fédérale de Lau-sanne

Redgrave P, Prescott TJ, Gurney K (1990) The basal ganglia: avertebrate solution to the selection problem?. Neuroscience89(4):1009–1023

Robertson B, Saitoh K, Ménard A, Grillner S (2006) Afferents of thelamprey optic tectum with special reference to the GABA input:combined tracing and immunohistochemical study. J Comp Neurol499(1):106–119

Roth G, Grunwald W (2000) Morphology, axonal projection pattern,and responses to optic nerve stimulation of thalamic neurons in thesalamander Plethodon jordani. J Comp Neurol 428(3):543–557

Saitoh K, Ménard A, Grillner S (2007) Tectal control of locomo-tion, steering, and eye movements in lamprey. J Neurophysiol97(4):3093–3108

Sánchez-Camacho C, Marín O, Ten Donkelaar HJ, GonzálezA. (2001a) Descending supraspinal pathways in amphibians: I. Adextran amine tracing study of their cells of origin. J Comp Neurol434(2):186–208

Sánchez-Camacho C, Marín O, Smeets WJ, Ten DonkelaarHJ, González A. (2001b) Descending supraspinal pathways inamphibians: II. Distribution and origin of the catecholaminergicinnervation of the spinal cord. J Comp Neurol 434(2):209–232

Sánchez-Camacho C, Martín O, Ten Donkelaar HJ, GonzálezA. (2002) Descending supraspinal pathways in amphibians: III.Development of descending projections to the spinal cord in Xeno-pus laevis with emphasis on the catecholaminergic inputs. J CompNeurol 446(1):11–24

Selset R, Doving KB (1980) Behaviour of mature anadromous char(Salmoalpinus L.) towards odorants produced by smolts of theirown population. ActaPhysiol Scand 108:113–122

Smetana R, Juvin L, Dubuc R, Alford S (2010) A parallel cholinergicbrainstem pathway for enhancing locomotor drive. Nat Neurosci13(6):731–738

Stephenson-Jones M, Samuelsson E, Ericsson J, Robertson B, Grill-ner S (2011) Evolutionary conservation of the basal ganglia asa common vertebrate mechanism for action selection. Curr Biol21(13):1081–1091

Takakusaki K, Saitoh K, Harada H, Kashiwayanagi M (2004) Role ofbasal ganglia-brainstem pathways in the control of motor behav-iors. Neurosci Res 50(2):137–151

Takakusaki K (2008) Forebrain control of locomotor behaviors. BrainRes Rev 57(1):192–198

Ullén F, Deliagina T, Orlovsky G, Grillner S (1995) Spatial orientationin the lamprey. II. Visual influence on orientation during locomo-tion and in the attached state. J Exp Biol 198(Pt 3):675–681

Ullén F, Deliagina TG, Orlovsky GN, Grillner S (1997) Visual path-ways for postural control and negative phototaxis in lamprey. JNeurophysiol 78(2):960–976

Want JJ , van der Nunes Cardozo JJ, Togt C (1992) GABAergicneurons and circuits in the pretectal nuclei and the accessory opticsystem of mammals. Prog Brain Res 90:283–305

van der Vesselkin NP, Ermakova TV, Repérant J, Kosareva AA, Kenig-fest NB (1980) The retinofugal and retinopetal systems in Lampe-tra fluviatilis. An experimental study using radioautographic andHRP methods. Brain Res 195(2):453–460

Viana Di Prisco G, Pearlstein E, Robitaille R, Dubuc R (1997) Roleof sensory-evoked NMDA plateau potentials in the initiation oflocomotion. Science 278:1122–1125

VianaDi Prisco G, Pearlstein E, Le Ray D, Robitaille R, DubucR (2000) A cellular mechanism for the transformation of a sen-sory input into a motor command. J Neurosci 20:8169–8176

123

512 Biol Cybern (2013) 107:497–512

Zelenin PV, Grillner S, Orlovsky GN, Deliagina TG. (2001) Hetero-geneity of the population of command neurons in the lamprey. JNeurosci 21(19):7793–7803

Zompa IC, Dubuc R (1998a) a) Diencephalic and mesencephalic pro-jections to rhombencephalic reticular nuclei in lampreys. BrainRes 802(1-2):27–54

Zompa IC, Dubuc R (1998b) Electrophysiological and neuropharma-cological study of tectoreticular pathways in lampreys. Brain Res804(2):238–252

123