available at https://github.com/nawrotlab/larvaworld input during positive and negative chemotaxis has been extensively studied ( 14, 16, 31, 32 ). There is general consensus that during appetitive chemotaxis increasing odor concentrations suppress pauses and turns leading to longer lasting bouts of activity, while turns are promoted during down gradient navigation. Turning is biased by the detection of minor concentration changes during lateral bending, a process described as active sensing ( 24, 28, 33 ). The opposite effect has been reported during aversive chemotactic behavior whether due to punishment or reward omission ( 34 ). Larva behavioral preference under conflicting olfactory stimuli has been established as a population-level metric in multiple settings from quantifying the formation of memory after associative learning ( 15–17, 35, 36 ) to detecting individual differences in genetically identical larva strains ( 37 ). This sensory-driven behavioral modulation does not seem to be affected by social cues, justifying the study of individual larva kinematics even during population-level experiments ( 38 ).

in behavior-based robotics allowing sequential calibration and modular integration of partial neuroscientific models under a common framework ( 39, 40 ).

We here propose a trilayer behavioral control architecture for Drosophila larva foraging as illustrated in Fig. 1 A. The bottom layer comprises three basic stereotyped behaviors: crawling, turning and feeding. Each is realized by a low-level local sensorimotor loop between motor effectors and sensory feedback - mainly proprioception and mechanoreception - and additionally gustatory input in the case of feeding. Integration of these basic behaviors gives rise to composite behaviors. Exploration in stimulus-free conditions requires crawling and turning while integration of all three basic elements allows exploitation of a nutritious substrate. The intermediate layer introduces salient sensory stimulation of different modalities therefore allows reactive behavior in the face of presented risks and opportunities. Modulation of exploratory behavior under appetitive or aversive stimulation enables coherent navigation along sensory gradients. Here, we consider odor-driven chemotaxis as a process of active sensing in which the larva constantly samples the odor concentration during lateral bending motions enabling odorscape navigation and odor source discovery. At the top layer, associative learning during prior experience biases the evaluation of recurring sensory stimuli resulting in modulation of innate odor valence and providing experience-dependent adaptation of navigation.

The proposed behavioral architecture is naive to the underlying neural mechanisms that generate the respective behaviors. We hereby populate the architecture with diverse configurations of candidate mechanisms and explore autonomous and integrated control exerted by each subsequent layer in increasingly-complex behavioral simulations from exploration to chemotaxis to adaptive learning in odor preference experiments.

Locomotory model. We simplify the larva to a two-segment body (Fig. 2 F). This abstraction allows describing the body state at any point in time through only three parameters. These are (i) position of a selected midpoint (Fig. 9), (ii) absolute orientation of the front segment θor, and (iii) bending angle θb between the front and rear segments. This approach is in line with the common practice of quantifying larva bending via a single angle between one anterior and one posterior vector, although the chosen vectors vary between study groups ( 14, 34, 41 ). Body dynamics as analyzed in the empirical data depends on instantaneous linear (Fig. 2 A-C) and angular velocities (Fig. 2 E-G) generated through crawling and bending respectively. Kinematic analysis on the locomotion of these simplified real larvae allows us to realistically calibrate bisegmental virtual larvae resembling the real ones in multiple spatial and temporal parameters (see Materials and Methods). For a demonstration of the larva-body bisegmental simplification see Video 1. The individual trajectory depicted in Fig. 2 can be seen at its full length in Video 3, while the slice depicted in the inset is shown in Video 4.

Locomotion of the bisegmental body is achieved via dynamic coupling of forward crawling and lateral body bending. Crawling moves the midpoint along the orientation vector. Bending reorients the front segment by rotation around the midpoint. Forward displacement restores θb back to 0 by gradually aligning the rear segment to the orientation axis.

We adopt the hypothesis that attributes alternating leftright bending to an underlying oscillatory neural process and compare the lateral oscillator model described by Wystrach et al. ( 24 ) to a simple sinusoidal oscillator. This model (turner) applies torque oscillations on the two-segment body, which acts as a restorative torsional spring (see Materials and Methods). The turner can be coupled to a second oscillator producing crawling strides (crawler) in the form of linear velocity oscillations as suggested by our kinematic analysis. Under the constraints of the subsumption architecture paradigm these effectors can be influenced by higher-order circuits only in a limited number of ways. Frequency modulation and stride initiation/cessation are the only possible top-down modulations on the crawler. Likewise, the turner can receive olfactory input 4 | affecting both the frequency and the amplitude of oscillation.

To quantify the contribution of each of the locomotory model’s components to replicating real larva kinematics we perform a broad model comparison study across diverse model configurations (Fig. 3). The behavioral metrics used for evaluation are structured in five categories covering angular and forward motion, reorientation, spatial dispersion and stridecycle structure. Overall, the bisegmental is superior to the single-segment body in capturing angular motion metrics, predominantly due to the decoupling of the bending velocity θ˙b from the orientation velocity θ˙or (in the single segment body these are considered identical). A realistic crawling oscillation allows much more accurate assessment of stride-cycle dependent metrics compared to non-oscillatory constant-speed forward motion (despite the lack of a velocity oscillation in the latter case, strides might still be detected due to noise). Finally, both oscillator coupling and crawl-intermittency contribute to better fitting of the empirical data. Our final model is represented by the far bottom right column, exhibiting the smallest error.

The detailed structure and optimal configuration for the basic layer of the behavioral architecture as suggested by model comparison is illustrated in Fig. 1 B, specifying the interplay between the oscillators. Two additional features are implemented based on our empirical analyses. First, the crawler and turner are phasically coupled such that turning is suppressed during a defined phase interval of the stride cycle, reflecting our finding in Fig. 2 H. Second, crawler-generated chains of concatenated strides (stridechains) are intermitted by brief pauses during which the crawler-induced interference is lifted, resulting in more acute turning events. The duration distribution of stridechains and pauses are fitted to the empirical data. The pipeline for model calibration is described in detail in Material and Methods. For a demonstration of the locomotory model in different configurations see Video 2. We note that in the current implementation effectors receive no sensory input from the environment. Sensory feedback is introduced only at the intermediate reactive layer (Fig. 1 A) via an olfactory sensor located at the front end of the body (head). All model parameters are shown in Table 1.

Inter-individual differences are crucial for achieving realistic population-level behavior. To capture variability across larva we computed three crawl-related parameters, body-length l, crawling frequency fc, and scaled stride displacement dstr across a population of 200 experimental larvae. The univariate and bivariate empirical distributions are illustrated in blue in Fig. 4. When generating virtual larva populations for the behavioral simulations we sample all three parameters from a multivariate Gaussian distribution fitted to the empirical data. This preserves the linear correlation structure. A generated set of the three parameters completely defines the crawling oscillation of an individual virtual larva. The bivariate projections are shown in red in Fig. 4 with parameters given in Table 1 in blue.

Simulation of behavioral experiments. The layered behavioral architecture in Fig.1 is exploited in modeling as it justifies sequential calibration and evaluation of subsequent behavioral layers from the bottom to the top, as for any specific behavior to be successfully realized, only integration up to a certain layer is required. In this section we simulate increasingly more complex behavioral experiments using virtual larva popula

Single-segment body

Two-segment body Uncoupled

Coupled

Uncoupled Intermittent Coupled Intermittent Oscillator coupling & Intermittency tions. Starting from stimulus-free exploration we advance to chemotactic navigation and finally to adaptive odor preference experiments. Individual virtual larvae behave independently of each other as they move through the spatial arena and odorscape ( 38 ).

Free exploration. Larvae explore a stimulus-free environment, dispersing in space from their starting positions. As there is no food nor any salient odor gradient, integration of crawling and bending behaviors is sufficient. We compared free exploration in populations of 200 virtual and real larvae respectively (Video 6). Statistical evaluation showed a good agreement of simulated and empirical data with respect to spatial dispersion of larvae from their initial position (Fig. 5 A,B), total distance traveled, time fraction allocated to crawling and number of performed strides (Fig. 5 C). For a demonstration of virtual and real larvae exploring a dish see Video 5 while for a comparative assessment of their dispersion dynamics see Video 6.

Chemotaxis. Chemotaxis describes the process of exploiting an odor gradient in space to locate an attractive or avoid a repelling odor source. An olfactory sensor (olfactor) placed at the front end of the virtual body enables active sensing during body bending and allows detection of concentration changes that modulate turning behavior accordingly (see Methods). To assess chemotactic behavior in our model we reconstruct the arena and odor landscape (odorscape, see Methods) of two behavioral experiments described in ( 14 ). In the first, larvae

are placed on the left side of the arena facing to the right. An appetitive odor source is placed on the right side. The virtual larvae navigate up the odor gradient approaching the source (Fig. 6 A), reproducing the experimental observation in Fig. 1 C in ( 14 ). In the second, both the odor source and the virtual larvae are placed at the center of the arena. The larvae perform localized exploration, generating trajectories across and around the odor source. (Fig. 6 B), again replicating the observation in Fig.1D in ( 14 ). Fig. 6 E and F show the average time-varying odor concentration encountered by the virtual larvae along their trajectories, replicating the estimations from real larval tracks in ( 14 ). Two sample simulations can be seen in Video 7.

Odor preference test. We simulate the odor preference paradigm as described in the Maggot Learning Manual ( 42 ). Larvae are placed at the center of a dish containing two odor sources in opposite sides and left to freely explore. The odor concentrations are Gaussian-shaped and overlapping, resulting in an odorscape of positive and/or negative opposing gradients. After 3 minutes the final situation is evaluated. The established population-level metric used is the olfactory preference index (PI), computed for the left odor as P Il = Nl−NNr where Nl and Nr is the number of larvae on the left and right side of the dish while N is the total number of larvae.

The extend of olfactory modulation on the turning behavior is determined by the odor-specific gain G (see Materials and Methods). As this is measured in arbitrary units, we first need to define a realistic value range that correlates with )4.8 m m l( th4.0 g n e ly d ob3.2 2.4 ) z (Hc2.0 f y c n euq1.6 e lfr w ra1.2 c )0.30 (trs pd0.24 e ts e ird0.18 ts 0.12 3.2 4.0 4.8 body-length l(mm) 1.2 1.6 2.0 crawl frequency fc(Hz) 2.4 0.12 0.18 0.24 stride step dstr(0.)30 Fig. 4. Parameters of individuality: empirical and fitted distributions. Diagonal: Histogram and kernel density estimates (KDE) for body-length l, crawling frequency fc and mean scaled displacement per stride dstr across a population of 200 larvae in the experimental dataset. Below: Bivariate projection of 3-dim. KDE outlined contours for each parameter pair. Above: Red ellipses represent the bivariate projections of the 3-dim. fitted Gaussian distributions at 0.5, 1, 2 and 3 standard deviations. In our model this Gaussian is used to sample a parameter set for each individual larva. The blue dots denote the empirically measured parameters. the behaviorally measured PI. We perform a parameter-space search independently varying the gain for left and right odors and measuring the resulting PI in simulations of 30 larvae. The results for a total of 252 gain combinations within a suitable range of G ∈ [−100, 100] are illustrated in Fig. 7 A. Simulation examples for one appettitive and one aversive odor are shown in Video 8.

In order to simulate larval group behavior in response to an associative learning paradigm we interface our behavioral simulation with the spiking mushroom body (MB) model introduced in ( 43 ) (Fig. 7 C). It implements a biologically realistic neural network model of the olfactory pathway according to detailed anatomical data using leaky integrate-and-fire neurons ( 44 ). The MB network undergoes associative plasticity at the synapses between the Kenyon cells and two MB output neurons as a result of concurrent stimulation with an odor and a reward signal. Both, odor and reward is simulated as spike train input to the receptor neurons and the reinforcement signalling dopaminergic neuron, respectively. The model employs two output neurons (MB+, MB-), representing a larger number of MB compartments associated with approach/avoidance learning respectively ( 45 ). The initially balanced firing rates between MB+ and MB- are skewed after learning and encode the acquired odor valence ( 46, 47 ) here defined as M Bout =

M B+ − M B− M B+ + M B− ∈ [−1, 1].

We first trained the MB model via a classical conditioning experiment where, in each conditioning trial, it experiences an odor (conditioned stimulus, CS+) in combination with a 6 | sugar reward during 5 min, following the standard training protocol in ( 42 ). 5 groups of 30 MB models undergo between 1 and 5 sequential conditioning trials ( 48 ). The resulting odor valence MBout from each MB model was converted to an odor gain G via a simple linear transformation and used to generate a virtual larva (Fig. 7 B). Each population of 30 larvae was then tested in an odor preference simulation. The larvae were placed on a dish in presence of the previously rewarded odor (CS+) and a neutral odor in opposite sides of the dish (Video 8), again following standard experimental procedures ( 42 ). To obtain robust results we replicated the experiment 100 times per population with a different random seed for a total of 600 simulations. The obtained preference indexes (PI) are illustrated in Fig. 7 D. The PI increase with increasing number of trials as well as its saturation resembles empirical observations ( 48 ). Note that the variability of the PI across the 100 simulations per condition is introduced solely by the behavioral simulation and resembles that seen across real experiments.

The current implementation only sequentially couples a trained MB model to be tested in a behavioral simulation. In the discussion we further elaborate on a possible extension featuring their closed-loop integration allowing for full behavioral simulations of both the training and the testing phase of the associative learning paradigm in a virtual environment.

B D premotor network downstream of the AL ( 16, 63 ). In the sensorimotor loop, descending pathways involving the LH control cessation of crawling, possibly triggering sharper reorientation when navigating down-gradient, facilitating chemotaxis ( 64 ). Finally, the internal homeostatic state (e.g. starvation vs. satiation) regulates behavior via neuromodulatory transmitter release at multiple levels, including AL, LH and MB ( 56 ). Existing computational models. Crawling and bending mechanisms have been successfully captured in previous computational models. CPG models of segmentally repeated paired excitatory and inhibitory (EI) neuronal rate units, standing for average EI population activity, can autonomously generate forward and backward crawling, possibly involving proprioceptive feedback ( 65, 66 ) and the contribution of the visceral-piston mechanism to the peristaltic cycle has been assessed in a biomechanical model ( 67 ). The hereby adopted, continuous oscillatory lateral bending process has been implemented as a pair of bilateral mutually inhibitory EI circuits ( 24 ). The idea has been elaborated in a neuromuscular model generating autonomous forward / backward crawling and turning as well as their interplay during free exploration in a 12-segment larva body by modeling segmental localized reflexes and substrate frictional forces and assuming empirically informed axial and transverse oscillatory frequencies ( 68 ). Chemotaxis has also been modeled computationally either in stochastic transition models assuming discrete behavioral states of crawling (runs) and turning (head-casts) ( 69 ) or by introducing olfactory modulatory input on the underlying locomotory circuit ( 24, 68 ).

We propose here that modeling locomotion as coupled intermittent oscillatory processes (Fig. 1 B) is adequate for generating realistic larva kinematics as these are captured via larva tracking. This locomotory model efficiently summarizes the underlying CPG activities into global linear and angular velocity oscillations ( 23, 26 ) and accurately reproduced a number of experimental observations. It is therefore well suited for the bottom layer of the proposed behavioral architecture (Fig. 1). The latter though allows for a modular approach, where the level of abstraction can be chosen independently for each individual module as exemplified by combining this locomotory model with a spiking neural network that captures plasticity in a central brain neuropile (MB) at the top layer. It follows that the proposed control architecture affords any substitution of the currently selected oscillator modules by one of the aforementioned more-detailed neuromechanical models if a higher degree of biological realism is pursued. Crawl-bend interference. Crawling includes mouth hook motion. Specifically, the first phase of a crawling stride consists of concurrent forward motion of head and tail segments, aided by a ’visceral pistoning’ mechanism that generates forward displacement of the gut. Subsequently, the mouth hooks anchor the head to the substrate so that the second phase of peristaltic motion can drag all other segments forward as well, completing the stride ( 23 ). Crawling and bending partially recruit the same effector neural circuitry and body musculature at least at the level of the thorax. Peristaltic motion during crawling includes sequential symmetric bilateral contraction of all segments while bending occurs due to asymmetric unilateral contraction of the thoracic segments. This partial effector overlap could result in interference between the two processes. Indeed here we report an increase of orientation velocity during a specific phase of the stride cycle (Fig. 2 H), synchronous to an increase of head forward velocity (Fig. 2 D). The latter coincides with the stride phase when the head is not anchored to the substrate and therefore free to move laterally. When applying attenuation of lateral bending outside a phase interval [ π2 , π] of the stride cycle we managed to accurately reproduce the empirical relation (Fig. 8 A).

A reasonable hypothesis would then be that the asymmetric thoracic contraction generating lateral bending is only possible while the head is not anchored to the substrate therefore during a specific phase interval of the stride cycle. We suggest that crawling phasically interferes with lateral bending because of these bodily constraints. A consequence of the proposed hypothesis is that the amplitude of turns generated during crawl-pauses is larger in comparison to those generated during crawling because during pauses the crawling interference to lateral bending is lifted. We postulate that it is exactly this phenomenon that dominates the description of larva exploration as a Levy-walk with non-overlapping straight runs and reorientation events, where the minor orientation changes taking place during crawling are neglected ( 10, 70 ). We note that in the implemented model the turner neural oscillation is not inhibited at all during crawling strides as we consider this merely a bodily interference, although the resulting torque might eventually not be applied to the body depending on the stride-cycle phase.

Behavioral intermittency. Larval locomotion is intermittent meaning that crawling runs are transiently intermitted by brief pauses. The spatial dynamics of these alternating states have been studied in the context of motion ecology. During free exploration, power-law distributed runs, in line with Levywalk theoretical models ( 10, 70 ) and diffusion-like kinematics have been reported ( 32 ) while the speed-curvature power-law relationship has been disputed ( 71–74 ). Regarding the temporal dynamics of intermittency, the duration distributions of activity and inactivity bouts captured via larva-tracking recordings have been studied. More specifically, the duration of inactivity bouts has been reported to follow a power-law while that of activity bouts a log-normal distribution ( 11 ), partly in line with findings in adult-fly studies ( 12, 13 ). In all these studies micro-movements like feeding and lateral bending could not be detected due to technical constraints. Thus, the reported inactivity and activity bouts can be regarded as crawl-pauses and crawl-runs respectively, the duration of the latter being equivalent to our discretized stridechain-length metric.

Our analysis reveals that both distributions are approximated best by log-normal distributions (Fig.8,B). The lognormal stridechain distribution is in line with previous findings, while the log-normal pause distribution diverges from a previously reported power-law ( 11 ). This might be attributed to the short, 3-minute duration of the recordings in the present study contrary to the long, up to one hour recordings used in the latter.

Computational models of behavioral intermittency are scarce. A recent study presented a simple binary-neuron model exhibiting state transitions between power-law and non A

-100 -100 GR -50 , n i a g r 0 o d o t h igR 50 100 power-law regimes via self-limiting neuronal avalanches and proposed a plausible underlying mechanism that explains initiation/cessation of crawling ( 11 ). In the current model we remain agnostic to the intermittency generative process and simply sample pause duration and stidechain length from the empirically fitted distributions. Further neuroanatomical evidence of the underlying circuits is needed to elucidate the neural mechanism of intermittent crawling.

Architecture extensions. The here adopted subsumption architecture paradigm is mainly used in behavior-based robotics ( 75 ) and is supported as a theoretical framework for the description of the nervous system of living organisms ( 39 ). According to this, neural control of behavior consists of nested sensorimotor loops where more stereotyped reflexive behaviors are autonomously generated by localized neural circuitry at a faster timescale without recruiting more centralized resources. Once there is need for more extensive integration, e.g. in order to react suitably to unexpected sensory stimulation, slower multisynaptic loops enable higher centers to modulate local circuits achieving more coordinated global behavioral control. In other words, the nervous system is considered a multilayered control architecture within which higher layers subsume their subordinate layers into comprehensive behavioral modes. The main idea governing this paradigm is that there are only limited degrees of freedom in which higher layers can influence the lower ones e.g. by initiation/cessation or acceleration/deceleration of their autonomous function.

The modularity of the proposed architecture facilitates further extensions capturing novel aspects of the body, the nervous system, the metabolism or the environment as shown in Fig.1, allowing for more refined or more complex behavioral patterns. We here summarize some plausible extensions. Feeding. Despite their central role in foraging, feeding mechanisms have not been modeled computationally nor has their integration with the exploratory circuitry in the context of substrate exploitation. Feeding behavior also consists of repetitive stereotypical movements of the head, mouth hook and ingestive muscles ( 27 ). Every cycle is an intertwined sensorimotor loop under instantaneous feedback from the environment and slower regulation by higher self-regulatory centers. In adult flies successive feeding movements are organized in intermittent feeding bouts, interspersed by locomotory or idle periods ( 76 ). Although detection of larval feeding motions in freeforaging conditions is technically difficult due to their small amplitude, it is reasonable to assume these are also structured in intermittent bouts. The frequency of this repetitive motion has been reported to vary at least from 1 to 2.5 Hz. Therefore feeding behavior can be implemented as a third oscillator (feeder). Top-down modulation affecting initiation/cessation and oscillation frequency can be assumed, similar to crawling while sensory feedback can be implemented as recurrent modulation depending on successful ingestion.

Considering oscillator coupling, crawling and feeding cycles partially compete for control of the same effectors as they both recruit the head and mouth muscles. It has been reported that the number of mouth hook motions over a given duration of foraging does not differ between rover and sitter larva phenotypes and is not correlated to the amount of food ingested, although rovers crawl more and feed less than sitters ( 30 ). Therefore an individual mouth hook motion can equally Sakagiannis et al. be part of either a crawling or a feeding cycle. The conclusion drawn is that individual crawling and feeding cycles are competing mutually exclusive behavioral motifs meaning crawling and feeding bouts can alternate but do not overlap. Conversely there is no empirical evidence on the potential coupling of feeding and bending motions. Integration of the feeder in the current locomotory model is illustrated in Fig.1 B. Multimodal sensory feedback. Sensory feedback from the environment can be extended to other modalities beyond olfaction. Mechanosensation can be implemented via additional touch sensors around the body contour. This will allow detection and behavioral modulation by conspecific contact (collisions) ( 77 ) and external mechanosensory stimulation driving startle/evasion (hunch/bend) ( 78 ) or navigation along wind gradients (anemotaxis) ( 79 ). Accordingly, temperature and light sensors could allow thermotaxis ( 80–82 ) and phototaxis ( 83 ). The respective environmental sensory gradients can be Gaussian as in the case of odorscapes or linear along a certain arena axis. Behavioral modulation can be either summed up across modalities or separately applied to the bending, crawling and feeding effectors. Integration of additional sensory modalities is illustrated in Fig.1 A.

Olfactory learning in closed loop behavioral simulations. We have shown an open-loop simulation of the classical conditioning paradigm (Fig. 7 B-D) reproducing a basic experimental result in the fruit fly larva ( 36 ). This modeling approach can be extended in multiple ways. First, the larva demonstrates a number of interesting learning abilities/features that require synaptic and circuit mechanisms such as differential conditioning ( 15, 16, 36 ), extinction learning ( 84 ), and relief learning ( 35, 45, 48, 85 ). Interfacing neural network simulations of these mechanisms with our behavioral model allows to directly compare virtual and empirical behavioral experiments, both for the typical group assays and for individual animals. Second, while information about odor concentration is provided via olfactory sensing, direct input from a feeder module could provide the reward stimulus that activates the dopaminergic pahtway required for synaptic plasticity the mushroom body ( 45, 86 ). This would further allow realistic foraging scenarios with food depletion and competition. Closing the loop from active sensing to associative memory formation and behavioral control requires to synchronize a (spiking) neural network at the adaptive layer with the sensory (reactive layer) and locomotor modules (basic layer). This will enable the simulation of virtual larva experiencing spatial and temporal dynamics in a virtual environment or on a robotic platform ( 87 ). Such model approaches will allow to test model hypotheses on sensorymotor integration and to infer predictions for experimental interventions such as optogenetic stimulation ( 45 ) or genetic manipulations ( 88–91 ).

Weber-Fechner law, meaning that ΔAo ∼ ln ΔC. We further add a decay term that slowly resets Ao back to 0. The rate of change is given by the equation : i

˙ X Gi · Ci

Ci A˙o = −Ao · co + with − 1 ≤ Ao ≤ 1 [4] where co = 1 is the olfactory decay coefficient, Gi is the gain for odor i and Ci the respective odor concentration. Perceived olfactory stimulation Ao modulates the turner activation At from its baseline value At = 20 within a suitable range Arange = [Atmin, Atmax] = [10, 40] :

At = A¯t + Ao Atlim − A¯t where

Atlim =

Atmax Atmin ifAo ≥ 0 ifAo ≤ 0 [5] [6]

˙ full length. Trajectory color tracks instantaneous forward v and angular θor velocities while tracking a midline point or the head respectively. Color code ranges from red (0 velocity) to green (maximum velocity).

˙ shown. Trajectory color tracks instantaneous forward v and angular θor velocities while tracking a midline point or the head respectively. Color code ranges from red (0 velocity) to green (maximum velocity).

Video 2. Locomotory model for Drosophila larva The function of the locomotory model in Fig. 1 B is illustrated by gradually integrating its 4 modules (crawler, turner, oscillator-coupling, crawling-intermittency). In each of the 6 videos the implemented modules are shown in the inset.

ACKNOWLEDGMENTS. This project was funded by the German Research Foundation (DFG) through a stipend for P.S. within the

shown for a population of 200 real (left) and virtual (right) larvae (Fig. 5 A).