Load compensation in targeted limb movements of an insect

Reaching movements made by primates and other vertebrates are compensated well against perturbations caused by external loads such as artificial force fields (e.g. Lackner and Dizio,1994; Shadmehr and Mussa-Ivaldi, 1994) or by internal loads such as joint interaction torques (e.g. Gribble and Ostry,1999; Koshland et al.,1991). There is considerable debate about whether muscle forces are encoded explicitly by the nervous system, or whether limb position is encoded and the movement arises as the result of changes in a neuronally specified `equilibrium position' (Feldman,1986). In insects, load compensation has been studied in the context of posture or gait control (e.g.Pearson, 1972;Bässler, 1977;Cruse, 1990;Dean, 1991;Noah et al., 2001), but not in the context of aimed limb movements where load sharing and other coordinating mechanisms between legs can be excluded. The goal of this paper is to determine whether such aimed movements in an insect are compensated against altered loading.

In walking stick insects the protraction (swing phase) movements of the rear legs are targeted at the position on the ground where the foot of the next anterior leg already has a foothold. Perturbations of the position of the anterior foot lead to adjustments of the touch-down position of the posterior leg (Cruse, 1979), and when the swing movement is resisted by an external force the targeting also remains accurate (Dean, 1984). When the swing movement of a rear leg is assisted by an external force, the velocity is increased only slightly. Taken together, Dean's data show that position is controlled for the endpoint of swing, but that velocity is controlled during the movement itself (reviewed in Dean and Cruse, 1986). Nevertheless, in walking, the control of any given leg is influenced by the actions of the others. For example, when the animal drags a load the velocity of swing movements is increased, presumably to permit a relatively longer stance phase(Foth and Graham, 1983). Such interactions between limbs can be excluded in the present study because (1)the load is added directly to the leg that is observed, (2) the leg moves to a target location that is independent of the postures of the other legs and (3)the aimed movement is carried out while the other legs remain still.

To test whether loading affects the performance of a targeted leg movement we analysed scratching movements made by a hind leg of a locust in response to tactile stimuli on the ipsilateral wing (see Matheson,1997,1998;Dürr and Matheson, 2003). We have previously shown that in unloaded conditions the pattern of leg movements used by locusts to move the distal end of their tibia towards different targets forms a continuum that is modified (shifted) by changes in target location on the wing surface. The location of the target signalled by wing exteroceptors is monitored throughout the movement so that re-targeting can occur to track a moving object(Matheson, 1998), but it is not known whether proprioceptive signals from the moving leg are similarly used to maintain accuracy.

To examine the effects of loading different leg joints, we loaded either a basal or a distal position on the femur or a distal location on the tibia. Loading the leg so that the rotational moment of inertia was increased by up to 11.6-fold had no detectable effect on any of the measured kinematic variables. We conclude that both position and velocity are controlled in this targeted movement, and that proprioceptive signals from receptors on the leg must be integrated throughout the movement with exteroceptive signals from the wing that signal target location.

The recording and analysis methods follow the protocols set out in detail in Dürr and Matheson(2003) and are summarised only briefly here for completeness.

Experiments were carried out on adult female desert locusts Schistocerca gregaria Forskål. They were tethered by a loop of fine wire that passed around their pronotum without obstructing movements of any of the legs and were suspended above a light (4.5 g) foam ball on which they stood or walked. The eyes were covered with solvent-free typists'correction fluid, and the tarsus of a hind leg was placed on a horizontal rod located at a fixed position (Fig. 1A) that corresponded to the anterior position used in Dürr and Matheson (2003). To elicit scratching movements of the hind leg, the ipsilateral forewing was touched gently with a fine paintbrush. Stimuli were applied to one of two different locations in pseudo-random order. These locations correspond to positions 2(anterior) and 5 (posterior) used in Dürr and Matheson(2003). The exact location of the stimulus and the resulting movements were measured from videotape as described below. A small metal bead of mass 142 mg was fixed using sticky wax(Lactona Surgident, Philadelphia, USA) to the hind leg at one of three different locations (a–c in Fig. 1A) to provide a load (the hind leg mass was approximately 138 mg). The added mass was chosen so that it approximately matched the leg mass. By adding this mass to different sites on the leg, the moment of inertia of a single joint was altered in a range from 1% to 1160% of the normal value. Each set of scratches elicited under loaded conditions was interspersed with a set of control scratches elicited under unloaded conditions. There was no evidence of a temporal change in scratching behaviour (i.e. learning) throughout the course of the experiments, which lasted for 6–30 h in total. Each animal was tested for periods of up to 90 min at a time separated by intervals of at least 90 min to avoid fatigue. Experiments were carried out at 22–24°C.

Animals were videotaped from the side using a calibrated CCD camera with a spatial resolution of 0.1 mm pixel^–1, allowing manual digitising accuracy of 0.5 mm, as determined from the standard deviation of all digitised points in repeated analyses of three sequences. The images were recorded on video tape, captured on a personal computer and deinterlaced to yield AVI files with a frame size of 768×576 pixels at a frame rate of 50 s^–1. A custom-written program (Borland Delphi) was used to access the AVI files and digitise the coordinates of up to 12 points per frame(Fig. 1B). The coordinates of the digitised points were output as a text file that was read into a second custom-written program for a variety of analyses (see below). The stimulus location and the position of the tarsus rod were digitised in the first frame of each trial. A further eight points were digitised in all frames to record movements of the stimulated wing and the ipsilateral hind leg. These were: (1)the front leg coxa, (2) the base of the wing on its midline, (3) the distal end of the wing on its midline, (4) the hind leg coxa, (5) the hind leg trochanter, (6) the distal end of the hind leg femur, (7) the distal end of the hind leg tibia and (8) the distal end of the second segment of the hind leg tarsus (Fig. 1A,B). Sequences were analysed from the first frame of tarsal movement until one of the following four events occurred: (1) the tarsus touched the ground, (2) the tarsus hit the stimulus brush, (3) the leg completed three complete cycles of movement or (4) the tarsus stopped moving for 120 ms.

453 loaded or unloaded (control) scratches were analysed in three animals,all of which expressed qualitatively the same behaviour. Two components of each response were distinguished: a short (200 ms) initial component, in which the trajectory of the tarsus was relatively straight, and a second cyclic component, in which the tarsus moved in repeated loops near the target (see e.g. Fig. 2). The average velocities of the initial and cyclic components were calculated separately. The cyclic component was further quantified by means of probability distributions, which described the likelihood that a particular part of the leg moved across any given point in the leg's work space (for derivation, seeDürr and Matheson,2003).

To assess the effect of load on the ability of locusts to make aimed movements of their hind legs, a mass of 142 mg was attached to different locations on the leg (Fig. 1A,B). This effectively doubled the total mass of the leg and, in particular, when the mass was added to the distal end of the tibia, it increased the mass of this segment from 19 to 161 mg (an 8.5-fold increase). The rotational moment of inertia for a joint depends on the square of the distance of the centre of mass from the axis of rotation. Each load position that we used therefore resulted in different changes in rotational moment of inertia at each joint (Table 1.) For example, loading the tibia caused an 11.6-fold increase in rotational inertia for the femur-tibia joint, but a 6.4-fold increase for the thorax–coxa joint (Table 1).

Locusts made aimed scratching movements of a hind leg (e.g.Fig. 1C) in response to tactile stimuli at one of two target sites on the dorsal surface of the ipsilateral forewing (the outermost surface when the wings are folded into the normal resting posture, Fig. 1A). The leg always started from the same position, which was defined by a small rod on which was placed the animal's tarsus (Fig. 1A).

Stimulation of the distal (posterior) site elicited movements in which the tarsus was lifted and moved posteriorly towards the target before making a variable number of cyclical movements in the vicinity of the target(Fig. 2Ai). Stimulation of the proximal (anterior) site elicited movements in which the tarsus was lifted and moved approximately vertically towards the target before again making cyclic movements largely dorsal to the animal(Fig. 2Aii). Only the first three cycles were analysed. For a detailed analysis of unloaded scratching movements and their natural variability, see Matheson(1997,1998) and Dürr and Matheson (2003).

Addition of a 142 mg load to the proximal femur, distal femur or distal tibia had no effect on the general form of scratching movements for either target site (compare Fig. 2Ai,ii with Bi,ii, Ci,ii and Di,ii). Slight variations in the movements illustrated in Fig. 2 fall well within the variance seen in unloaded scratches, as we go on to demonstrate in the following sections.

To examine the movements in more detail we analysed separately the initial 200 ms of movement that formed the outgoing trajectory and the remaining part of the movement during which the tarsus followed a cyclical path. The use of 200 ms as a cut-off criterion is justified quantitatively in Dürr and Matheson (2003). It is the mean duration of the outgoing trajectory. When the leg was unloaded the median direction of movement in the first 200 ms was 112° for scratches elicited by stimulation of the anterior site (black vector inFig. 3A), and 134° for scratches elicited by stimulation of the posterior site (black vector inFig. 3B). Loading the leg had no significant effect on initial movement direction (coloured vectors and curved lines in Fig. 3A,B:Dunnett's two sided t-test for each treatment versuscontrol, 453 scratches from three animals pooled; all values of P>0.05). Each animal was also analysed separately. For movements to the anterior site, loading the distal femur caused a significant increase in the angle of movement for one animal (Dunnett's t-test, P=0.005, N=28 unloaded, nine loaded scratches), but a decrease in the second (P=0.047, N=36, 15), and had no effect in the third (P=0.089, N=57, 10). For movements to the posterior site loading the distal femur caused a significant reduction in the angle of initial movement in one animal (P=0.005, N=43,12). No other load condition caused any significant difference in initial movement direction in any animal (all values of P>0.05).

In the unloaded condition, 89% of scratches aimed at the anterior target site had three or more cyclical loops (Fig. 4A). Loading the proximal or the distal femur led to a small but non-significant reduction in the relative frequency of occurrence of scratches with three loops (black bars in Fig. 4A: χ²=14.7, 9 d.f., P>0.05). Only 40%of unloaded movements aimed at the posterior target had three loops, with most of the remainder having one or two loops(Fig. 4B). Once again, loading the leg caused a small but non-significant reduction in the proportion of scratches with three loops when compared to the unloaded condition (black bars in Fig. 4B:χ ²=16.9, 9 d.f., P>0.05).

When the leg is unloaded the part of the leg that is most reliably aimed at the stimulus, irrespective of stimulus location, is the distal end of the tibia (Dürr and Matheson,2003). To test whether locusts change this aiming strategy when the limb is loaded we determined the point on the leg that came closest to the target under each loading condition. For unloaded movements made in response to stimulation of the posterior target site, the distal end of the tibia and the base of the tarsus were, on average, aimed most accurately(Fig. 5Ai). The minimum values of closest distance (i.e. highest accuracy) reached 0 mm only for the distal end of the tibia and tarsus, indicating that the proximal tibia and femur could not reach the target (white line inFig. 5Ai). For movements made in response to stimulation of the anterior target the distal end of the tibia was aimed most accurately (Fig. 5Aii). The morphology of the hind leg means that when a locust rotates the femur forward to move its tarsus near the anterior target (but not the posterior one), a point on the femur may also cross the target (see e.g. Figs 1,2Cii). This results in a local minimum in the measures of closest distance for the femur inFig. 5Aii. It is important to note that the femoral minimum is the inevitable result of forward rotation of the femur that is required to bring the distal tibia and tarsus close to the target. The tibia crossed the stimulus site repeatedly during the cyclic part of grooming, whereas the femur was typically rotated forwards so that it passed the target only once and was then held anterior to the target throughout the response (see Fig. 1C). This indicates that, although the femur can pass over proximal targets, it is not specifically aimed at them.

Loading the leg had no effect on the aimed point for any of the three load sites or either of the targets (Fig. 5Bi,ii–Di,ii). The absolute measures of accuracy for each point along the leg were similar to the corresponding values for unloaded movements (compare Fig. 5Bi,ii–Di,ii with correspondingFig. 5Ai,ii), and the overall pattern of accuracy was the same in all cases. For unloaded scratching movements we have shown previously that a point 4 mm from the distal end of the tibia is the most consistently aimed part of the leg(Dürr and Matheson, 2003,asterisk in their fig. 2). Because load has no effect on movement strategy, we used the same point for all further analyses in the present paper (asterisks inFig. 5Ai,ii).

There was a weak but significant positive correlation between the initial velocity (over the first 200 ms) and the velocity during the cyclical part of the movement in all three animals analysed separately and in the pooled data(Pearson correlation coefficient r=0.285, P<0.001, N=414; note that not all responses included a cyclical component). The median velocity of movement of the distal end of the tibia during the first 200 ms was unaffected by load for either the anterior stimulus (open bars in Fig. 6A) or the posterior stimulus (grey bars in Fig. 6A). This was also the case for each animal analysed separately(data not shown).

The velocity of movement in the cyclical part of the response was unaffected by load on the proximal femur for scratches aimed at either stimulus site (Fig. 6B). In contrast, loading the distal femur or tibia caused an increase in the median velocities for movements made in response to the posterior stimulus (two extreme right grey bars in Fig. 6B; 43% and 48% increases, respectively).

Two measures were used to investigate the effect of loading on the cyclical part of the movement. First, the point of closest approach to the target was determined to obtain a measure of performance accuracy. Second, the average`movement distribution' was calculated to describe the region of the leg's workspace covered during the responses.

For unloaded movements to the posterior target, the mean point of closest approach was 5.0±2.7 mm (mean ± s.d.) anterior to the target (Fig. 7). For unloaded movements to the anterior target, the point of closest approach was 1.9±2.5 mm posterior to the target. Loading the leg had no effect on the point of closest approach in any experimental situation.

To assess the part of the leg's workspace that is most likely to be scratched, we calculated the likelihood that the distal part of the tibia crossed each point in space between two subsequent video frames, given a particular stimulus:load condition. This was done by counting the number of times the distal tibia crossed a given location in the leg's workspace during a single trial, and standardising the resulting distribution to a volume of 1. Averaging these two-dimensional distributions resulted in distributions of the average likelihood of observing a movement across any point in the workspace.Fig. 8 illustrates this as grey level maps in body coordinates (for details of the calculation, seeDürr and Matheson,2003).

In the unloaded condition, the cyclical parts of scratching movements aimed towards the posterior target (Fig. 8Ai)differed from those aimed at the anterior target(Fig. 8Aii). The probability distribution for posterior scratches was centred anterior to the target site,extending both ventrally and dorsally above and below the wing(Fig. 8Ai). The centre of each distribution was characterised by the `centre of gravity' (white circles inFig. 8) and by the `most likely point' (white squares in Fig. 8). For unloaded movements aimed at the posterior target, both measures lay approximately 8 mm anterior to the target near the ventral(leading) edge of the wing. The probability distribution for anterior scratches peaked approximately 7 mm posterior and dorsal to the corresponding target (Fig. 8Aii). The centre of gravity was a further 4 mm posterior and dorsal, reflecting the `tail' of the distribution.

Bayes' theorem (see e.g. Quinn and Keough, 2002) was used to test whether the probability distributions resulting from movements to different targets or made under different load conditions could be distinguished from each other. This was done by estimating the minimum number of video frames in a scratching sequence that was required to determine statistically which experimental condition had given rise to the observed behaviour. For any given load condition, the probability distributions for movements made in response to stimulation of the anterior target site (Fig. 8Aii–Dii) could be distinguished easily from those made in response to stimulation of the posterior target(Fig. 8Ai–Di; seeTable 2).

Loading the leg had no significant effect on the probability distributions for movements made to the posterior target(Table 3,Fig. 8Bi–Di). For movements aimed at the anterior target, loading the leg increased slightly the density of the tail of the distribution, reflecting the terminal downward movement of the trajectory as the tarsus returned to the ground. Accordingly the centre of gravity (but not the most likely point) shifted ventrally(Fig. 8Bii–Dii). Despite this subtle change, the distributions for movements to the anterior target were also statistically indistinguishable(Table 3). Observation of more than 450 video frames (which is approximately equivalent to the duration of 10 scratches), would have been required to reliably detect the different loading conditions. In addition to these analyses of the pooled data from three animals, we examined each animal separately. None of the load conditions had any significant effect on the probability distributions in any of the animals(all P>0.15), with the single exception that loading the distal femur caused a marginally significant shift in one of the animals(P≤0.04).

The marked similarity of the probability distributions obtained under different load conditions becomes even more apparent when examining the pattern of iso-density contours that delimit different proportions of the total volume (Fig. 9). Contours that contained only the top 10% of volume (i.e. delimited the peaks of the distributions) overlapped each other almost completely, especially for anterior scratches (Fig. 9, top row). Contours for successively greater proportions of the total volume (25,50, 75 and 90%) revealed an approximately symmetrical pattern of expansion for posterior scratches, irrespective of load condition(Fig. 9, left column). For anterior scratches the 50% contours revealed the posterior–ventral tail of the distributions, which was more pronounced for loaded trials (e.g.Fig. 9, 50%, right). The 90%contour revealed a slight anterior dorsal shift and a less pronounced tail when the proximal femur was loaded (Fig. 9, 90%, right; red line).

We have shown that applying a substantial load to one hind leg of an insect does not impair the animal's ability to make well coordinated, aimed movements of that leg towards targets on its wing. The general form of the movements is the same under all conditions (Figs2,3,4), as is the initial velocity of movement (Fig. 6). Animals aim the same part of their distal tibia at the target regardless of loading(Fig. 5), and they remain equally accurate (Fig. 7). Loading has no effect on the region of the workspace that is scratched (Figs8,9). The mass on the distal femur or tibia applied a load to the coxal levator muscles during scratches aimed at both stimulus locations on the wing, but also loaded depressor muscles for movements to the anterior site as soon as the load passed through the vertical. The mass on the tibia loaded the tibial extensor muscle during the outward trajectory aimed at the distal target, but loaded the flexor muscle for the initial movement to the proximal target. The same mass alternately loaded flexors and extensors during the cyclical part of scratching. The load was substantial, representing an 11.6-fold increase in the rotational moment of inertia of the tibial segment, so its lack of effect on the velocity, accuracy and trajectory of leg movement indicates clearly that leg proprioceptive inputs are integrated in the motor networks that generate aimed scratching. In the following discussion we rule out the alternative possibility that leg stiffness is sufficiently high that an unchanging equilibrium point control mechanism could produce the same trajectory irrespective of loading.

To move the tarsus along a particular trajectory to a target requires computation of motor commands that will act to rotate the individual leg joints against their specific internal and external loads. This requires that the motor signals are represented in a coordinate system that encodes the magnitude of joint rotations, torques or muscle actions – i.e. it must be an intrinsic coordinate system, as opposed to the extrinsic system used by sensory receptors to encode the spatial location of the tarsus and target. In theory, an equilibrium point mechanism could permit the planning and execution of the trajectory of a coordinated reaching movement, even for a redundant manipulator like the locust leg (Feldman and Levin, 1995). Such a control mechanism requires high joint stiffness, however, particularly if the trajectory is to remain similar after changing the moments of inertia. Passive rotations of the locust femur–tibia joint that are observed during ongoing leg movements(Berkowitz and Laurent, 1996)indicate that stiffness of this joint is low in the absence of external load. During scratching the femur–tibia joint can extend with considerable force in the absence of extensor motor activity, at least in part as a result of energy transfer from the movements of more proximal joints(Berkowitz and Laurent, 1996). In other words, the stiffness of the unloaded femur–tibia joint is insufficient to stabilise the joint against the tibia's intrinsic moment of inertia. Only if proprioceptive information was to adapt joint stiffness to the additional load could joint stiffness explain the load compensation that we describe.

Joint stiffness is affected by three active mechanisms: cocontraction of antagonist muscles, neuromuscular inhibition and neuromodulation. The timing of activity of common inhibitory neurones during active movements must contribute to the low leg joint stiffness by downregulating tonic contraction forces of antagonist muscles at each joint(Wolf, 1990). Insect leg muscles are controlled not only by their excitatory and inhibitory motor neurones, but also by the activity of specific neuromodulatory neurones[`dorsal unpaired median' (DUM) neurones; reviewed inBurrows, 1996]. These leg DUM neurones can affect muscle tonus and force production(Hoyle, 1978) so, if their activity is modified as a result of loading, they might also contribute to load compensation. Nevertheless, none of these mechanisms could participate in load compensation in the absence of proprioceptive feedback, so we rule out the possibility that a pure equilibrium point control mechanism underlies the load compensation we observed.

Our conclusion concurs with Dean's claim(Dean, 1992) that if stick insects use an equilibrium point mechanism to control targeted stepping movements, then they must also use proprioceptive signals. Indeed, most current versions of the equilibrium point hypothesis accept that joint interaction torques are not automatically compensated for by joint stiffness,but that internal models of the limb's properties must be calibrated by sensory feedback (e.g. Hollerbach and Flash, 1982; Shadmehr and Mussa-Ivaldi, 1994; Ghez and Sainberg, 1995). Human arm stiffness is too low to permit pure equilibrium point control (Gomi and Kawato,1996,1997), and subjects who lack proprioception due to large-fibre neuropathy show marked errors in limb trajectory (Gordon et al.,1995). In spinal frogs, where loading of a hind leg has little or no effect on targeting accuracy (Schotland and Rymer, 1993), proprioception has similarly marked influences on initial trajectory direction, path straightness, knee joint velocity and overall accuracy (Kargo and Giszter,2000).

Proprioceptive feedback is clearly essential for the generation of normal motor patterns in most rhythmic motor systems (reviewed inPearson and Ramirez, 1997). In insects the angle, velocity and acceleration of each leg joint are monitored by internal chordotonal organs (Hofmann et al., 1985; Matheson and Field,1990; reviewed in Field and Matheson, 1998), the receptors of which have powerful reflex effects on motor neurones of the same joint(Field and Burrows, 1982) and of other joints of the same leg (Hess and Büschges, 1999). In the locust, information about hind leg joint angles excites particular spiking local interneurones that are also excited by exteroceptive inputs that signal a touch on the wings(Matheson, 2002). These interneurones are likely to be key elements in the generation of aimed leg movements and in compensation against load perturbations.

Increased load affects muscle forces that are monitored by multipolar tension receptors, which also have intra- and interjoint effects(Theophilidis and Burns, 1979;Matheson and Field, 1995). Load-induced strain in the exoskeleton is sensed by campaniform sensilla(Spinola and Chapman, 1975),which are generally located near joints(Pringle, 1938). Those on the coxal and trochanteral segments near the base of the leg are key elements in load-compensating reactions during walking, when the leg is in contact with the ground (e.g. Zill et al.,1999; Noah et al.,2001). These campaniform sensilla are generally inactive when the leg is off the ground (as during scratching). Nevertheless, loading the leg as we did is likely to have caused activation of these receptors, when muscles were acting against increased rotational inertias to accelerate and decelerate each leg segment throughout the scratch. Whether the resultant feedback pathways alone (e.g. Newland and Emptage,1996) could provide the complete control of movement velocity and accuracy that we see is unknown. Signals from campaniform sensilla converge onto leg motor neurones along with position-dependent signals from other leg sense organs (Schmitz and Stein,2000), and the leg motor neurone output is modulated by non-linear interactions between the two modalities.

In humans, loading a limb results in complex EMG responses, the gain of which depends on the simultaneous angular motion of other limb joints(Lacquaniti and Soechting,1986). This implies that the spinal cord integrates proprioceptive inputs from several muscles. Some neurones in the motor cortex are sensitive to loading at either the shoulder or the elbow, or respond to loads at both locations (Cabel et al., 2001). In insects, sensory influences on leg movements have been demonstrated in a wide range of experiments (for reviews, seeBurrows, 1996;Cruse et al., 1984;Bässler and Büschges,1998), and there is widespread convergence of proprioceptive information from different leg segments onto leg motor neurones. This holds for joint position- and movement-sensitive chordotonal organs (Hess and Büschges, 1997,1999), load-sensitive campaniform sensilla (Zill et al.,1981), and muscle tension receptors(Matheson and Field, 1995). In a walking stick insect, the protraction movement of a posterior leg is aimed at the position where the next most anterior leg already has a foothold(Cruse, 1979). Loading the posterior leg has little effect on either the accuracy or velocity of its movement (Dean, 1984). In walking, however, the control system driving these movements is constrained by the need to coordinate the posterior leg with the weight-bearing legs so as to maintain a stable gait. Our data show complete compensation for loading during a limb-targeting behaviour in which the aimed limb is not subject to such coordinating influences. The control of movement velocity is therefore not a consequence of the requirement for interleg coordination, but instead appears to be a general feature of insect leg movements across species and in different behaviours. In compensating for increased leg loading, locusts did not trade-off accuracy for velocity, nor did they switch to an alternative movement strategy. Because trajectories do not change when the leg is loaded,and targeting is strongly dependent on the exteroceptive cue of a tactile stimulus (Dürr and Matheson,2003), we conclude that load compensation in locusts must be due to a high-gain position control loop.

Our previous work shows that in locusts, unloaded scratching movements elicited by stimuli to different places on the wing form a fine-grained continuum (Dürr and Matheson,2003), suggesting that a single underlying motor pattern is modulated in a continuous way by a somatosensory representation of the wing surface. The present results show that locusts, like vertebrates, also integrate leg proprioceptive information into an internal body model that is used to control aimed limb movements.

This work was supported by a BBSRC Advanced Research Fellowship, a Royal Society Research Grant and a Newton Trust grant to T.M., and a FIF2 grant of the University of Bielefeld to V.D. We thank H. Cruse, M. Gebhardt and P. L. Newland for reading a draft of the manuscript, A. Goncalves DaSilva for help with digitising, and Steve Rogers for advice on the statistics.