Mechanisms of octopus arm search behavior without visual feedback

The octopus coordinates multiple, highly flexible arms with the support of a complex distributed nervous system. The octopus’s suckers, staggered along each arm, are employed in a wide range of behaviors. Many of these behaviors, such as foraging in visually occluded spaces, are executed under conditions of limited or absent visual feedback. In coordinating unseen limbs with seemingly infinite degrees of freedom across a variety of adaptive behaviors, the octopus appears to have solved a significant control problem facing the field of soft-bodied robotics. To study the strategies that the octopus uses to find and capture prey within unseen spaces, we designed and 3D printed visually occluded foraging tasks and tracked arm motion as the octopus attempted to find and retrieve a food reward. By varying the location of the food reward within these tasks, we can characterize how the arms and suckers adapt to their environment to find and capture prey. We compared these results to simulated experimental conditions performed by a model octopus arm to isolate the primary mechanisms driving our experimental observations. We found that the octopus relies on a contact-based search strategy that emerges from local sucker coordination to simplify the control of its soft, highly flexible limbs.


Introduction
The octopus employs its eight highly flexible arms across a range of behaviors, including foraging, exploration, manipulation, and locomotion.Each arm has hundreds of suckers staggered along its length.These suckers bear a complex chemotactile system (Graziadei 1964, 1965, Graziadei and Gagne 1976, van Giesen et al 2020) and are the primary appendages used by the octopus to interact with its environment (Packard et al 1988).
Most of the octopus's nervous system is distributed into its arms and suckers (Young 1971) where it takes the form of an axial nerve cord extending down the center of each arm.This nerve cord (using the terms recommended by Richter et al 2010) consists of two bundles of axonal tracts running dorsally to a medullary nerve cord, which consists of a neuropil surrounded by a cortex of cell bodies.The dorsal axonal tracts include pathways connecting the arms and brain as well as locally interconnecting pathways (Graziadei 1971, Young 1971).Nerve roots projecting ventrally from the nerve cord transmit information between the nerve cord and suckers, while the dorsal nerve roots transmit information between the nerve cord and the arm musculature (Gutfreund et al 2006).Efferent motor pathways from the brain do not target specific locations of the arm and instead broadly innervate a large pool of motor neurons along the nerve cord.Where a given behavior is activated is likely triggered by combining this broadly innervating signal with local sensory input specifying the site of activation (Zullo et al 2019).This organization of sensory and motor circuitry in the arm results in much of the arm's behavioral repertoire remaining intact when isolated from the brain (Rowell 1963, Altman 1968, Sumbre et al 2001, Gutfreund et al 2006, Zullo et al 2011).
The brain and arms communicate over a limited neural bandwidth (Young 1965) and a great deal of information within the arms does not ascend to the brain.A number of observations (Wells and Wells 1957, Wells 1964, Rowell 1966) suggest that proprioceptive information conveying the configuration of the arms and suckers is not represented in the brain.Wells describes mechanical information acquired by the suckers as translating into a frequency signal which lacks information about the exact pattern and orientation of encountered surface features.Chemosensory information has been shown to ascend through multiple levels of integration starting within the sucker epithelium, and the number of chemosensory channels accordingly shows a dramatic reduction as they ascend to the nerve cord (Graziadei and Gagne 1976).However, octopuses can learn to distinguish different chemicals and different concentrations of seawater (Wells 1963), suggesting that enough chemical information ascends to the brain to learn these distinctions.van Giesen et al (2020) identified a novel class of contact-dependent chemotactile receptors and showed that these receptors, in addition to distinct chemoreceptors and mechanoreceptors, are capable of flexible strategies of information coding.While proprioceptive representation of the arms appears to be absent from the brain, the octopus has demonstrated an ability to use visual feedback to guide its arm to the location of a reward without mechanical or chemical cues (Gutnick et al 2011).
Possibly due to the lack of sensory feedback from the arms, reaching behavior relies on a feedforward activation mechanism (Gutfreund et al 1998, Sumbre et al 2001), which gives the arm a ballistic kinematic profile as it extends toward its target.This suggests that the brain only takes into account the vertical and horizontal angle (yaw and pitch) of a target and activates reaching behavior in this direction without any further modification of the behavior following its activation.Where this information is sufficient for the retrieval of a reward, the octopus shows an improved performance over time without visual feedback (Gutnick et al 2020).Beyond these limited parameters that the brain is able to control and recall, the suckers and their local nervous system likely provides the necessary sensory-motor feedback to adapt behavior to the environment.
In tasks where visual feedback is absent and feedforward activation is inadequate for the retrieval of a reward, the octopus must rely primarily on the sensory system of the suckers in the generation and modification of behavior (Gutnick et al 2020).These conditions also remove any visual information that could possibly indicate arm configuration (Gutnick et al 2011).Characterizing the strategies that are employed by the arms and suckers to search for and retrieve a reward in the absence of visual feedback can therefore help resolve how the chemotactile system is used to control the arm's extreme flexibility.Here, we investigate the strategies used by the Pacific red octopus (Octopus rubescens) when searching for and retrieving a food reward from a visually occluded foraging task space.
Previous studies have described a behavior within the arm referred to here as sucker recruitment, which is commonly reported as a grasping mechanism.During this behavior, a sufficiently strong stimulus applied to a sucker results in the neighboring suckers bending toward the source of this stimulus.This same effect can then be elicited in these neighbors and cause a propagating wave of recruitment down the arm.An early observation of this effect was made by Ten Cate (1928), who demonstrated that the pathway for this signal projects through the arm nerve cord and that this pathway remains active in isolated arms.Altman (1968) reported that following stimulation, recruitment appeared first in distal suckers before proximal suckers.When recruitment occurred in response to an object, suckers serially adhered to the object upon contact.Altman likewise reported that stronger simulation elicited recruitment in more distant suckers.
While the exact circuitry making up the recruitment pathway is not known, studies have found that the activated sucker signals the neuropil through the ventral nerve roots, where it propagates along a polysynaptic pathway as it signals the target arm and sucker musculature through the dorsal and ventral roots.The activation of this musculature orients the neighboring sucker toward the stimulus, which can repeat this effect and propagate the signal onward.Along this pathway additional sources of motor and sensory information are considered which could either excite or suppress this signal.Provided that enough suckers can propagate the signal or the activation is sufficiently powerful, recruitment can propagate over great lengths of the arm (Rowell 1966, Altman 1968, Gutfreund et al 2006, Zullo et al 2011).
This recruitment mechanism can serve as an effective strategy for object handling and prey capture, and could likewise serve as a potential foraging strategy by allowing the arm to adapt to the shape of surfaces in search of prey hidden within unseen crevices.This form of surface conformation could further provide a means for the octopus to simplify control of its movement by providing constraints on the arm's vast degrees of freedom.The reliance on contact would be consistent with a force control (rather than position or velocity control) strategy by the octopus.
We therefore predict that if the octopus reaches its arm through a narrow entrance into an open space like that of our task, sucker contact with the entrance will initiate a recruitment signal which will attract the arm toward the surfaces of the space shared by the entrance.Additionally, we predict that the arm will preferentially search along concave edges and vertices where multiple surfaces meet.Such contours could serve to confine the arm's range of possible configurations and thereby simplify behavior.
To substantiate the role of sucker recruitment in generating our experimental observations, we created a simple computational model of the octopus arm and subjected it to simulated experimental conditions that matched our 3D-printed experimental task space.This model included a recruitment mechanism that activated between serially joined segments.This mechanism operated in parallel to random motion generated within the joints.This random motion was amplified at the proximal-most joint of the model, which corresponded to the section of the octopus's arm reaching through the task entrance.

Methods
Subjects included eight Pacific red octopuses (O.rubescens) collected using SCUBA from the Salish Sea under an approved scientific collection permit through the Washington Department of Fish and Wildlife (18-347 & 20-148).Subjects were held in recirculating aquarium systems containing artificial seawater or flow-through natural seawater systems.Because subjects were wild caught, they were kept under an artificial light cycle that roughly corresponded to natural daylight hours.Tank sizes ranged in volume from 3 to 40 gallons (roughly 10-150 l) depending on the size of the animal.During the course of the study, subjects were offered food (scallop or shrimp meat) equal to 1%-2% of their mass daily as the reward for the task.This was done to prevent satiation from affecting engagement with the task.All animals were provided with regular enrichment during the course of the study.All experiments were carried out in accordance with a protocol (4356-02) approved by the University of Washington Institutional Animal Care and Use Committee.
Figure 1 outlines our experimental design.The foraging task and its interior task space were created using computer-aided design (CAD) software and 3D-printed with white polylactic acid (PLA) filament.The task space was shaped as a simple box with a small half-cylinder centered on the top face leading to the entrance.Compared to the similar paradigm of Gutnick et al (2020) who used a narrow Y maze, this task space allowed the arm more unrestricted movement.The task was secured to the side of a glass or transparent acrylic tank using magnets, such that the task space was visible to a camera outside the tank but not to the octopus.At the top of the task was a circular entrance which opened into the center of the half-cylinder.The entrance was large enough for a single arm to fit through.This entrance was at the bottom of a larger opening designed to guide the octopus's arms toward the entrance.A light was used to illuminate the task from behind to maximize the contrast between the task space features and the octopus's arm.An acrylic barrier prevented the octopus from creating shadow artifacts by reaching between the task and the light.Prior to the start of trials, the octopuses were trained to approach the entrance of the task for a food reward while the interior was blocked off.
The task was printed in two sizes depending on the weight of the subject (which ranged from 38 to 125 g).Octopuses over 80 g were assigned the large task and octopuses under 60 g were assigned the small task, which was 80% the size of the large task.Octopuses between 60 and 80 g were randomly assigned to either version.For one condition ('Middle'; n = 4 males), the food reward was secured in the center of the box's largest face, and in the other condition ('Side'), the reward was secured to either the left (n = 2; one male and one female) or right wall (n = 2; one male and one female) of the box close to the concave edge between the wall and the box's largest face.Video recording failed for the first trial of one subject in the side condition, so this trial was not included in the analysis.The reward was equally distant from the entrance in both conditions.
The octopuses were given as many opportunities as needed to retrieve the reward six times from the task.Each of these six trials included the total time the octopus spent searching the task space until the reward was found, and this metric was used to define performance.An attempt was defined as the time between when an arm entered the task space and when it either found the reward or left unsuccessfully.Trials often consisted of multiple failed attempts before a successful attempt was made, and the duration of the failed attempts leading to success was factored into performance.Trials lasted over multiple days, but because larger subjects were given more food according to their weight, they were generally able to complete more trials in one day.An example of a successful attempt for the middle and side task can be seen in supplemental video 2 and supplemental video 3, respectively.Arm behavior was recorded at 250 FPS with a CMOS camera (Imaging Source, model DMK 37BUX287) and custom-written software in LabVIEW (National Instruments).Arms were tracked using DeepLabCut pose estimation software (Mathis et al 2018) (see supplemental figure 1), and analysis was performed using custom routines written in Python.

Arm tracking
A set of 861 frames were labeled as the training dataset for DeepLabCut.A maximum of 50 labels were placed in order along the dorsal edge of the arm opposite the suckers.Labels started with '1' at the arm tip and continued until the entire visible arm was labeled.A separate network was trained to track task features including the task entrance.Effective implementation of DeepLabCut depends on labeling unique features.However, the only consistently unique feature of the octopus arm was the tip.A consecutive, evenly spaced labeling pattern along the arm resulted in frequent occurrences of large sections of the arm not being found.We therefore introduced variability into the labeling pattern by pseudorandomly spacing apart and skipping labels.This led to improved tracking, which could be attributed to the more flexible labeling method allowing the network to generalize more effectively when tracking an ambiguous feature.
A spline was interpolated from the label coordinates of the tracked videos to define the arm's location.This tracking method led to inconsistencies between frames that, when quantified with the high video framerate, produced a great deal of noise.This was solved in our case by applying a temporal median filter to the spline coordinates, but subsampling frames would have also been sufficient.These coordinates were used to characterize kinematic properties of the arm including speed, curvature, and proximity to the outer walls of the task space.Using tracked task space features, the arm splines were centered and scaled to calculate arm occupancy across all trials.

Unity model
The computational model was developed using the Unity game engine with routines written in C#.This model consisted of a chain of 17 rigid body segments connected by joints with three degrees of freedom (yaw, pitch, and roll).Each segment measured 6.5 mm in length and included a capsule shaped collider (with a radius of 2.5 mm) to define the segment's collision boundaries with the task features and reward.Based on calculations by Yekutieli et al (2005), the mass of each segment was assigned a mass of 0.3 g and a drag coefficient of 0.1.To simulate sucker adhesion, torque was applied to segments toward any contacted surfaces.Sucker recruitment was simulated as torque additionally applied to the neighboring segments toward these surfaces.The intensity of this torque was chosen qualitatively so that recruitment realistically balanced the additional forces acting on the segments.The axis upon which this torque was applied was calculated from the cross product of the surface's normal (perpendicular) vector and the segment's forward vector (toward the joint shared with its distal neighbor).Motion was generated by randomly reconfiguring the arm's joints along their yaw and pitch between 0 • -180 • for the proximal-most joint and 0 • -36 • for the others.A torque was applied to each joint in the direction of this target pose after a randomly selected delay between two and three seconds.This torque equaled 3 Nm at the proximal most joint and decreased linearly along the length of the model to 0 at the distalmost segment.
The task CAD file was imported into Unity and the model was placed as though it was reaching into the task with its proximal-most segment fixed just outside the entrance.A boundary was added to the open side of the taskspace to simulate the side of the tank.
Two simulation versions were used to compare the model to our behavioral data.To compare performance of the model with the observed arm data, the model began each trial in a straight vertical pose extending down from the entrance.A simulated trial was completed when the arm made contact with the reward, which approximated the size and shape of the food rewards from the experiments.One thousand trials were run to simulate each condition.An example of a simulated trial for the middle task and both versions of the side task are shown in supplemental video 4.
To compare occupancy, the model was given 1000 s to search the task space without a reward to find.This was done instead of using repeated simulations because the model began each simulation in the same vertical position.Using one continuous simulation therefore prevented any artifact occupancy that resulted from repeatedly starting in this position.

Results
Performance for the observed data was assessed based on the time taken to reach the reward over the six trials and between the two conditions (time to success, figure 2).There was no improvement in performance across the six trials for either condition, but there was notably consistently better performance on the side task across trials.When trial averages were compared between conditions, the octopuses performed significantly better on the side task (Mann-Whitney U, p < 0.05).
Three kinematic measures were used to characterize arm behavior during the octopus's attempts to find the food reward: arm segment speed, curvature, and wall proximity (figure 3).All three of these measures showed a recurring behavioral profile represented by a distally oriented wave of deceleration, curvature and movement toward the wall, suggesting that the proximal arm leads the initiation of this behavior.The wave appeared to be a result of a distally propagating wave of sucker recruitment initiated and continuously regenerated by sucker contact with the task space features and resulted in the arm conforming to the shape of these features.This kinematic profile is also consistent with Altman's (1968) observation that sucker adhesion occurs with recruitment.This preference for the arm to conform to the outer walls of the task space was likewise reflected in arm occupancy calculated across animals and trials (figure 4), showing the arms spending most of their time conforming to the sides of the task during the experiment.
Using the model, we ran 1000 simulations for each condition (500 for each side).We then compared the performance distribution for each simulated condition to our observed performance data using a Kolmogorov-Smirnov (K-S) test.Results suggest that the observed and simulated arms share a distribution for both the middle task (p = 0.38) and side task (p = 0.20), showing that the model accurately simulated performance.
We ran a single simulation for 1000 s to characterize the preferred occupancy of the model (see figure 4).Like the experimentally-observed arms, the model preferred to conform to the shape of the task space features, which suggests that the model accurately simulated arm occupancy.

Discussion
To describe the control problem that the octopus faces, we will introduce the concept 'configuration space' used in the field of robotics to represent the multidimensional space of all possible configurations of a robot, where each axis represents a degree of freedom of the robot's joint angles.The more joints a robot has, the higher the dimensionality of its configuration space.With the octopus arm able to bend at any continuous point along its length in any direction, the octopus's space of possible configurations is immeasurably large (Kennedy et al 2020).This is all the more striking given that the octopus employs its arms across a wide range of behaviors.By using limbs with effectively infinite degrees of freedom across a variety of adaptive behaviors, the octopus has solved   Here we investigated the strategies that the octopus uses to search for prey without visual feedback, which is often the case when foraging within crevices, under rocks, or in conditions of low visibility.In these cases the role of vision is removed, and without a central representation of arm configuration, the octopus must rely primarily on the chemotactile feedback received by the suckers and the behaviors locally controlled within the arm's nerve cord.
To identify the mechanisms underlying these behaviors, our approach is to program these mechanisms into our computational model, subject the model to simulated experimental conditions, then compare the model's behavior to that of the octopus arm.The more accurately the mechanism is replicated within the model and the larger role the mechanism plays in the arm's behavior, the more closely the model would reasonably match the experimental data.
The prevalence of sucker recruitment in behavior with and without a connection to the brain made it a compelling mechanism to investigate within this modeling framework.As this mechanism alone would cause the model to stick motionless against a surface, we paired it with random motion generated within the segments.This motion was amplified in the proximal-most joint to simulate the proximally-led, distally oriented recruitment signal observed from the kinematic data.
Together, these mechanisms resulted in a pattern of the proximal segment 'casting' the arm in a random direction, where the rest of the segments then serially conformed to the task space features.Between these larger casting motions, the random motion of the distal segments caused the arm to move laterally across the surface.
The performance and occupancy of our experimental results suggest that the arms preferentially conformed to the shape of the task space's surface features, particularly along its concave edges, and that this behavior led to greater performance when the reward was found on the side of the task space.
To isolate sucker recruitment as the underlying mechanism of this behavior, we compared the performance distributions of our experimental data with those of our simulations.This comparison indicated that the observed and simulated arm share a distribution for both the middle and side task.
While the random motion generated a performance and occupancy profile that resembled the observed arms, the octopus likely employs a more systematic search strategy.For example, when conforming to a surface, the octopus arm's lateral movement is most likely not equally probable in both directions.Also, rather than generating larger proximal movements within a random interval of time, this kind of behavior probably occurs after the novelty of a searched area is depleted.It is reasonable that these two factors are not as limiting when the reward was found in the middle, which may have accounted for the closer resemblance in performance between the model and the observed arm for the middle task.The possibility of a systematic search strategy by the arm is interesting as it implies the existence of a mechanism allowing the octopus to distinguish areas that have been searched from those that have not.Our model further differs from reality in that it does not take into account chemical cues.Chemical cues could inform a more directed search by the arm when near the reward, however, these were not simulated within our model.Instead, if the model comes near the reward, it is not guided by these local sensory cues.Meanwhile, the octopus's suckers are densely innervated with chemoreceptors that may be picking up both the chemical cues from a nearby food source and its relative direction (Chase and Wells 1986, Walderon et al 2011, Fouke and Rhodes 2020).Chemical cues could also aid in search by indicating the areas the arm has already investigated (Wells 1963).
We attribute the difference in performance between the two observed conditions to a few primary factors.As the octopus searched the task space, the one definite location where the arm was making contact with surface features was where it was reaching through the entrance, and from this point of contact we believe a strong recruitment signal was being sent distally.Though the middle reward was secured to a surface, this surface was separated from the entrance by a concave edge oriented orthogonally to the direction of the reward relative to the entrance.The proximal-to-distal recruitment signal, likely guided by this edge, oriented the arm away from the middle and toward the side.The recurring kinematic profiles appearing in the arm speed, curvature, and wall proximity time plots reflected the prevalence of this behavior.Because of these factors, the most likely way for the arm to find the middle reward was when proximal-to-distal recruitment caused the distal arm to sweep past the reward while switching sides of the task space.
Distal-to-proximal recruitment was observed, though it usually occurred only after distal suckers found the food reward.Free proximal suckers were then 'reeled in' toward the reward (see supplemental video 2).This was not included in the analysis primarily because this behavior was difficult to track with DeepLabCut.This behavior indicates that while distal suckers are limited in their ability to capture and manipulate prey because of their size, they may serve as scouts by locating prey then recruiting proximal suckers to capture it.
The extent to which morphology and neuroanatomical organization is conserved across units of the sucker and its adjacent length of the arm's nerve cord and musculature has not been fully characterized.It is therefore informative to ascertain the degree to which the octopus arm's behavior was simulated in a model whose segments were controlled by identical routines.
While it has been shown that in locomotion, separate functional roles tend to be adopted by different lengths of the arms (Mather 1998, Hooper 2015, Levy et al 2015, Levy and Hochner 2017), it seems that in the context of search behavior (and possibly other behaviors employing similar recruitment patterns), sucker-arm units may be functionally identical.
By reflexively conforming its arm to the shape of the surrounding surfaces to guide its movement, these surfaces can confine the enormous degrees of freedom of the arms to a more manageable range of configurations.Serving as a lower dimensional reference during behavior, surfaces, especially sharp concave features, appear to act as coastlines in the arm's configuration space and are perhaps used like coastal navigation is used by ships (Roy et al 1999).
This behavioral strategy based on contact with the environment is interesting, as in the field of robotics collision with environmental surfaces is generally avoided.For the soft-bodied octopus arm, rather than being avoided, collision appears to be exploited as a control strategy.However, unlike the simple architecture of our task, the surface features of the octopus's environment are convoluted and complex.The next step for this paradigm will therefore be to investigate how this strategy is employed with tasks more closely representing this kind of surface complexity.
Our investigation includes a number of additional limitations.These include the use of too few trials to assess learning capabilities of the octopus (compared to studies where learning was evident, e.g.Wells and Wells 1957, Wells 1963, Gutnick et al 2020) the lack of available females (since all animals were wild-caught, sex of animals included in the study was not under our control), and our inability to distinguish between arms with this task design (with the exception of the third right arm in males, due to the presence of the hectocotylus).The questions that we are unable to address due to this latter limitation are potential arm preference, individual arm performance, and individual arm improvement.With the results presented by Bowers et al (2021), suggesting a possible peripheral mechanism for memory encoding within the arms of the dwarf cuttlefish Sepia bandensis, these are clearly interesting and viable research questions that this paradigm should seek to address moving forward.Efforts for future work will therefore be made to modify task design to distinguish arms.
While arm tracking with DeepLabCut was effective for our purposes, we discourage this approach for future studies of octopus arm behavior.In addition to the large amount of customization needed to effectively track the arm, the lack of large datasets for octopus behavior limited our tracking capabilities to two dimensional coordinates.With the central role of the suckers in the arm's behavior, investigations should pursue tracking methods that can identify individual sucker position and orientation.
The paradigm we used for this investigation involved using CAD software to design and 3D print a task space where we could investigate arm behavior in specifically designed conditions.These methods for using 3D printed tasks for studying arm behavior, like those used by Buresch et al (2022), present a number of advantages.As we describe here, task CAD files can be used to simulate experimental conditions for computational models where the validity of these models can be assessed.Additionally, given the precision of 3D printers, the same task can be printed in multiple locations and used for multiple species, ensuring replicability between investigations.
Sucker recruitment presents a simple peripheral behavioral mechanism that can lead to a number of adaptive advantages.By orienting suckers toward relevant stimuli, recruitment signals can act as an effective capture strategy.Multiple suckers can be recruited in immobilizing and handling prey where a single sucker may be insufficient.Given the minimal bandwidth between the brain and arms and the level of abstraction of mechanical information from the sucker disks (Wells and Wells 1957), any one sucker's ability to communicate with the brain is evidently limited.By recruiting their neighbors toward a relevant stimulus, the representation of a stimulus within the brain can be compounded through the collective sensory fields and afferent pathways of recruited suckers.Sucker recruitment can thereby serve as an adaptive, locally controlled sensory filter.Additionally, as supported by our results, sucker recruitment serves as an effective search strategy by allowing the arm to conform to the shape of surface features, and a mechanism by which the octopus can exploit these surface features as a means to shape its arms with minimal feedback to the brain.This mechanism allows the octopus to delegate the decision of arm shape to the arms themselves and chemotactile makeup of the environment, alleviating the brain of the computations necessary to reconfigure limbs with infinite degrees of freedom.

Figure 1 .
Figure 1.Experimental design.(a) Middle and (b) side task CAD design with dimensions of both large and small task versions displayed.(c) Experimental setup showing position of the camera, task, and light.(d) Octopus during a task attempt.(e) Model arm during a simulated middle task.

Figure 2 .
Figure 2. Performance as time to success.(a) Performance over six trials (as median and interquartile range), and regression showing no improvement for either condition.(b) Performance compared between conditions as trial averages, showing significantly better performance for the side task (Mann-Whitney U; p < 0.05).(c) Middle cumulative probability distributions of observed and simulated performance, showing shared distribution (K-S; p = 0.38).The maximum observed and simulated trial duration lasted 610.48 s and 916.56 s, respectively.(d) Side cumulative probability distributions of observed and simulated performance, also showing shared distributions (K-S; p = 0.20).* Video recording failed during the first trial of one subject in the side condition, so this trial was not included in the analysis.

Figure 3 .
Figure 3.(b) Example timeplots representing arm curvature (top), speed (middle), and wall proximity (bottom) during a middle attempt (shown in supplemental video 1) using the large version of the task.Horizontal axes represent time.Vertical axes represent the length along the arm (mm) as though the arm were laid along the axis with the base at 0 mm and the arm tip at 110 mm.(a) and (c) Recurring kinematic profile of a distally propagating wave of curvature, deceleration, and movement toward the wall.Note that curvature is represented here by the angle between segments in degrees, such that lower values represent greater curvature.

Figure 4 .
Figure 4. Most commonly occupied areas of the task for (left, 'Observed') the observed arms as proportion of total time across all trials and all animals (equally weighted) and for (right, 'Simulated') the model arm over 1000 s of a continuous simulation (without a reward).Both the arms of the observed data and model in the simulated data were represented by a single pixel-width line to calculate these plots.