A therian mammal with sprawling kinematics? Gait and 3D forelimb X-ray motion analysis in tamanduas

Gait and posture in quadrupedal tetrapods has often been broadly categorised into ‘sprawling’ with laterally oriented proximal limb elements (stylopodia) and abducted elbows/knees, and ‘erect’ with proximal limb elements oriented and moved within a parasagittal plane (Gregory, 1910, 1912; Gatesy, 1991; Gambaryan and Kielan-Jaworowska, 1997; Gambaryan and Kuznetsov, 2013; Lin et al., 2019a). This sprawling versus erect locomotor paradigm has received considerable attention in the literature of comparative tetrapod biomechanics (Gambaryan and Kielan-Jaworowska, 1997; Blob, 2001; Reilly et al., 2006; Regnault and Pierce, 2018; Granatosky et al., 2019; Fahn-Lai et al., 2020; Usherwood and Granatosky, 2020; Brocklehurst et al., 2021; Jones et al., 2021).

Sprawling locomotion is further characterised by pronounced lateral bending of the spine, and long-axis rotation (LAR) and retraction of the stylopodia, with the foot placed lateral of the shoulder and hip parasagittal planes (Barclay, 1946; Hildebrand, 1974; Edwards, 1977; Ritter, 1992; Ashley-Ross, 1994; Irschick and Jayne, 1999; Baier and Gatesy, 2013; Biewener and Patek, 2018; Baier et al., 2018; Nyakatura et al., 2019). In contrast, in parasagittal locomotion, the manus (autopodium) is moved within the shoulder-to-hip parasagittal plane and the elbows are oriented caudally (Hildebrand, 1974; Gambaryan and Kielan-Jaworowska, 1997; Biewener and Patek, 2018). Limb movements within the parasagittal plane can be more erect or more crouched (Riskin et al., 2016). Forward propulsion of parasagittal mammals is mainly achieved by retraction of the most proximal limb elements, which are the mobilised scapula of the pectoral girdle and the femur, in the forelimbs and hindlimbs, respectively (see Fischer and Blickhan, 2006). It has been shown that forelimb kinematic patterns of quadrupedal placental and marsupial (i.e. therian) species are highly preserved across clades (Kuznetsov, 1985; Fischer, 1994; Fischer et al., 2002), attributed to biomechanical benefits related to the typical zig-zag limb configuration (reviewed by Fischer and Blickhan, 2006). It becomes obvious that in terms of posture a spectrum can be imagined with perfect sprawling and parasagittal limb movements on opposite ends. Concordantly, various degrees of stylopodial abduction of sprawling species (Gatesy, 1991; Blob and Biewener, 1999; Blob, 2001; Nyakatura et al., 2019) and an increasingly recognised variety of stylopodial abduction angles from a parasagittal plane in (non-cursorial) mammals (Jenkins, 1971; Jenkins and Camazine, 1977; Schmidt and Fischer, 2000; Panyutina and Makarov, 2022) have been reported.

Strikingly high degrees of elbow abduction are a prominent feature in locomotion of the two extant species of Tamandua (Taylor, 1978), resembling a sprawling elbow positioning during stance phase (Fig. 1A; Fig. S1). This led us to investigate kinematically whether tamanduas incorporate aspects into their locomotion that are generally associated with a sprawling gait and would thus further populate the middle ground between the opposite ends of the spectrum described above.

The genus Tamandua, comprising two allopatric but morphological similar species (Hayssen, 2011; Navarrete and Ortega, 2011), belongs to the Xenarthra, that also include armadillos and sloths (Gibb et al., 2016). Together with the giant anteater (Myrmecophaga tridactyla) and the silky anteater (Cyclopes didactylus), tamanduas form the family Myrmecophagidae, characterised by extra joints in the lumbar vertebrae and the lack of teeth (Gaudin, 2003; Gaudin and McDonald, 2008; Oliver et al., 2016; Hautier et al., 2018). Tamanduas use terrestrial and arboreal habitats, feeding almost exclusively on termites and ants (Rodrigues et al., 2008; Brown, 2011; Hayssen, 2011; Navarrete and Ortega, 2011; Ferreira-Cardoso et al., 2020). Accordingly, the genus exhibits anatomical characteristics related to their arboreal locomotion, such as a prehensile tail (Hayssen, 2011; Navarrete and Ortega, 2011), whereas the forelimb displays numerous adaptations for hook and pull digging (Taylor, 1978, 1985; Kley and Kearney, 2007). While the noticeable abducted elbow of tamanduas (Taylor, 1978) and the placement of the lateral side of the hand on the ground during locomotion have been described (Pocock, 1924; Taylor, 1978; Polania-Guzmán and Vélez-García, 2019), no detailed description on locomotion and forelimb kinematics that underlie these prominent features is available.

We here describe locomotion of Tamandua tetradactyla qualitatively and quantify kinematic parameters. We first quantify metric gait parameters of tamanduas while walking on flat ground and on differently inclined branches using high-speed standard light video recordings. Second, we provide three-dimensional (3D) in vivo motion data of the 7th cervical vertebra (C7), scapula, humerus, ulna, and radius of the forelimb during treadmill locomotion and during balancing on a horizontal branch, to account for the arboreal lifestyle. The data were obtained using X-ray motion analysis and scientific rotoscoping (SR) (Gatesy et al., 2010), a marker-less variant of X-ray reconstruction of moving morphology (XROMM) (Brainerd et al., 2010). Finally, the kinematic data were further compared with data for both sprawling and non-sprawling tetrapods, to identify possible locomotory specialisations in tamanduas, generally associated with a sprawling gait.

Overall, four individuals of Tamandua tetradactyla (Linneus 1758) were available for this study. For the gait analysis, the animals were filmed with standard light cameras locomoting on differing support types and inclinations. Two adult individuals (one female, one male, unknown body mass) housed at the Tierpark Berlin, Germany, were filmed at 50 frames s⁻¹, both during locomotion on level ground and during balancing on wooden branches (diameter ca. 7 cm) within their cage. Movement was encouraged with food rewards offered by an animal keeper. The branches had inclinations of 0, 45 and 90 deg and the animals walked both up and down the supports.

Moreover, standard light videos (250 frames s⁻¹) obtained during the 3D kinematic experiments (see below) of two different adult individuals acquired from the Zoo Dortmund, Germany (female, 7.1 kg; male, 6.7 kg) were analysed. This was done in order to obtain additional recorded trials for the analysis of footfall patterns at varying speeds on a treadmill, a horizontal branch and a vertical branch (Table S1), thus increasing the number of total strides and range of exhibited speeds. The gait analysis comprised a total of 78 trials with 139 strides performed during 7 different tasks (ground locomotion, treadmill locomotion, and balancing on branches at inclinations of 0, +45, −45, +90 and −90 deg), henceforth called support types. Ground locomotion and treadmill locomotion were pooled and referred to as ground locomotion, as both experimental setups resulted in steady-state locomotion on a level and flat surface.

All videos were processed using ImageJ version 1.51 (Schneider et al., 2012) and all analyses and graphs generated from our data were done in R version 3.5.2 (http://www.R-project.org/).

When using inferential statistics to draw conclusions from comparisons between the support types, random factors must be included in the model to account for the non-independence of the data points. Non-independence was unavoidably introduced by including more than one stride per individual (animals were recorded on multiple support types and multiple times on the same support type, and multiple strides per trials were used), and potentially as a result of the close relatedness of the individuals available for the study. Given the limited number of specimens, the effect of sex could also not be reliably accounted for as a fixed factor. We judge the number of data points that we were able to gather in the time granted to us in the zoo and in the laboratory to be insufficient to account for these factors properly and have therefore refrained from testing for statistical significance of the observed differences between the support types. Still, as these issues are common for such experimental studies with exotic animals, we judge descriptive statistics of the data to be insightful for the assessment of gait adjustments of tamanduas on differing support types. Additionally, linear regression analyses and the coefficient of determination were used to describe trends in the data.

Gait classification follows the convention based on footfall patterns and each foot's relative ground contact duration introduced by Hildebrand (1966, 1976) and revised by Cartmill et al. (2002). We adopted the terminology of Cartmill et al. (2002). Diagonality (D) was determined by the percentage of the stride interval the footfall of a forefoot lags that of the ipsilateral hindfoot. Two states can be determined with D: a lateral sequence gait (LS) (with D<0.5) and a diagonal sequence gait (DS) (with D>0.5). At 0.25<D<0.75 the phases of diagonal forelimbs and hindlimbs are closely related (diagonal couplets). Therefore, a gait with 0.25<D<0.5 is classified as a lateral sequence/diagonal couplets (LSDC) gait. A gait with 0.5<D<0.75 is thus classified as a diagonal sequence/diagonal couplets (DSDC) gait. With 0<D<0.25 and 0.75<D<1, a footfall pattern resembles a pace (ipsilateral forelimb and hindlimb phases are closely related, i.e. lateral couplets). A gait with 0<D<0.25 is classified as a lateral sequence/lateral couplets (LSLC) gait. Footfall patterns in which D equals exactly 0, 0.25, 0.5 and 0.75, called pace, LS single-foot, trot and DS single-foot, respectively, are considered as idealised gait types (Cartmill et al., 2002) and rarely occur in nature. According to Hildebrand's convention, we used the duty factor (S; percentage of stride interval that each foot is on the ground) to differentiate walks (S>0.5) from runs (S<0.5). The duty factor index (DFI) is defined as the hindfoot duty factor expressed as percentage of the forelimb duty factor. A DFI below 100 indicates that the forelimb contacts the ground longer than the hindlimb and vice versa for a DFI above 100.

The two adult individuals of T. tetradactyla from the Zoo Dortmund, Germany, were temporarily housed solitarily at the Institute of Veterinary Anatomy of the Leipzig University, Germany, where the in vivo experiments for SR were conducted. In brief, we followed the workflow presented by Gatesy et al. (2010): (1) acquisition of biplanar X-ray videos focusing on the forelimb of moving tamanduas (Fig. 1A); (2) undistortion, calibration and contrast enhancement of X-ray videos and separation into image stacks; (3) acquisition of polygon mesh surface models of the bones of interest; (4) construction of a digital marionette of the bone elements (Fig. 2); (5) reconstruction of the experimental setup in animation software Autodesk^® Maya^® (version 2016, Autodesk, Inc., San Rafael, CA, USA) using calibration data to position virtual X-ray sources projecting the obtained X-ray videos onto virtual screens; (6) registration of the digital marionette via posing the model against the X-ray images on each virtual screen; (7) extraction of motion data and analysis. Each step is described in more detail below/in the next section. All procedures and animal care were carried out in accordance with the animal welfare regulations of the state of Saxony, Germany (registration no.: TVV 16/17).

Step 1: normal light high-speed videos (Optronics Cam Record CR600x2, Kehl, Germany) and X-ray videos were taken synchronously from two X-ray sources (Philips Medio 65 CP-H X-ray Generator, Amsterdam, The Netherlands and CPI Indico IQ Ultra High Frequency X-ray Generator, Ottawa, ON, Canada) and two corresponding image intensifiers (Philips Type BX 3i-2123, Amsterdam, The Netherlands) at a distance of 1.4 m (Fig. 1). The two X-ray sources were placed with an inter-beam angle of approximately 60 deg and operated in continuous mode at a voltage of 42 to 70 kV and a conduction current of 6 mA. All videos were recorded at 250 frames s⁻¹, a shutter speed ranging from 1 to 2 ms, a resolution of 1024×1024 pixels, and a maximal video length of 8 s.

Two locomotor tasks are investigated in this study: walking on a treadmill and balancing on a horizontal branch. The treadmill was regulated to match the speed chosen by the animals (see Results). The balancing branch was a birch tree trunk of 2 m length and with a diameter of ∼4.5 cm, placed horizontally at 1.8 m height. The animals walked voluntarily on the treadmill and were motivated by a spoonful of yoghurt in balancing trials. The two X-ray sources were positioned to allow recordings from a fronto-lateral and a caudo-lateral perspective.

Step 2: before each set of trials, X-ray images were obtained of a metal grating fixed flush to the faceplate of the image intensifiers and of calibration objects for virtual camera positions (see Gatesy et al., 2010; Knörlein et al., 2016). The X-ray videos were converted to image stack .tiff files, undistorted, and calibrated in XMALAB version 1.5.1 (Knörlein et al., 2016). Contrast was enhanced in Adobe^® Photoshop^® CC version 2015.1 (Adobe^® Systems Inc., San Jose, CA, USA).

Step 3: 3D surface mesh models of the scapula, humerus, ulna and radius were created using photogrammetry from collection material (specimen ID: ZMB 77026) of the Museum für Naturkunde, Berlin, Germany. The models were assembled using Agisoft^® Metashape^® (version 1.5.3, Agisoft LLC, St Petersburg, Russia). The surface model of C7 was created in Amira (version 6.0.0, Amira software, Thermo Fischer Scientific, Berlin, Germany) (see Stalling et al., 2005) from a micro-computed tomography (µCT) scan (specimen ID: ZMB- MAM- 91288, Museum für Naturkunde, Berlin, Germany).

Step 4: the digital marionette was constructed following protocols described by Gatesy et al. (2010) and Baier and Gatesy (2013) (Fig. 2). The centre of rotation (COR) of the joints was estimated by creating a sphere which was matched to approximate the curvature of the articular surface, placing the COR at the sphere's centre (e.g. Regnault and Pierce, 2018; Nyakatura et al., 2019; Richards et al., 2021). On top of the hierarchical ‘chain’ was an artificially created walking surface, followed by C7, scapula, humerus, and ulna and radius on the same hierarchical level. A detailed summary of the creation of the digital marionette is provided in Table 1. The right limb was created by mirroring in Maya^®. Based on the X-ray images, inner-joint spacing and size of bone models was adjusted for each individual, resulting in two slightly different sized model setups. Besides size, no obvious anatomical differences became apparent between museum specimens and the individuals of the SR analysis.

Maya^® uses a right-handed coordinate system and describes motion using Euler angles with a hierarchical rotation order (see Robertson and Gordon, 2004; Brainerd et al., 2010; Bonnan et al., 2016; Andrada et al., 2017), resulting in six degrees of freedom (DOF) data (rotations and translations about three axes). Following previous studies (Brainerd et al., 2010, 2016; Baier and Gatesy, 2013; Bonnan et al., 2016), the rotation order was set to ZYX corresponding to protraction/retraction (rotation about Z-axis), followed by abduction/adduction (Y-axis) and LAR (X-axis). In this way, rotation about the Z-axis also moved the Y- and X-axes as well as the bone model; rotation of the Y-axis also rotated the X-axis and the model; rotation of the X-axis only moved the model (see Robertson and Gordon, 2004). The model setup served as a reference position and did not reflect naturally occurring bone positions (Nyakatura and Fischer, 2010a; Bonnan et al., 2016). All reported rotations are reported in relation to the hierarchically higher ordered (proximal) element. Positive and negative values correspond to deviations from the zero position (model setup), described further in Table 1.

Anatomical rotations are defined as follows. Protraction is the cranial displacement of the distal end of a segment and retraction is its caudal displacement. Abduction is the displacement of the distal end of each element away from the medial plane, adduction is towards the medial plane. External LAR refers to a counter-clockwise rotation, around the long axis of the left limb's elements (in dorsal view). In the right limb, it refers to a clockwise rotation accordingly. Naturally, these rotations are produced in the proximal joint of the elements. For example, humeral protraction corresponds to extension in the shoulder joint, ulnar protraction corresponds to flexion in the humeroulnar joint. However, in this study rotations are described in three DOF, while language use of rotations in what is usually considered a hinge joint is typically restricted to a single DOF, namely flexion/extension. To avoid confusion, we therefore decided to report element rotations (protraction/retraction, abduction/adduction, LAR) instead of rotations within the joints.

Step 5: virtual camera positions obtained by calibration (step 2) were imported into the 3D workspace in Maya^®. The model setup was then placed at the intersection of the field of view of both Maya cameras (Fig. 1B).

Step 6: following the hierarchical chain in top-down order, each element was oriented to match both X-ray projections. C7 and the scapula were both translated and rotated, as these are not connected by natural joints to any other element of our model (Fujiwara, 2018). All other elements were rotated around the COR defined in Maya^®. Resolution was not sufficient to account for inner-joint translations. Therefore, translations were only measured for C7 and the scapular fulcrum, resulting in only 3 DOF in the other limb joints. In the treadmill setup, only the elements of the left limb were analysed; in the balancing setups, the right limb was also rotoscoped. Obtained values for abduction/adduction (Y-axes) and LAR (X-axes) of the right limb were subsequently multiplied by −1 to account for the mirrored model setup. A summary of the number of trials per experimental setup is provided in Table 2. To assess the accuracy of the procedure, a single stride was rotoscoped 5 times with at least 1 week between SR sessions. The results were evaluated by comparing the standard deviations (s.d.) for each element and each DOF (presented in Table S2). Further, the positions of the distal radius served as a proxy to determine placement of the hands in relation to each other (see Results), to compare ground and balancing hand positioning.

Step 7: translational and rotational values were extracted and are reported in centimetres (cm) and degrees (deg), respectively. If the preliminary data produced outliers, we used this information to reassess model positions at specific frames.

The field of view was too small to record a complete stride [e.g. touchdown (TD)–lift-off (LO)–TD] in the balancing trials. Therefore, we conducted a separate analysis of stance phase and swing phase in all SR analyses to ensure comparability between treadmill and balancing trials. Further, the results of the balancing trials needed to be assembled according to available sections of a stride in the individual X-ray recordings, resulting in highly idealised composed stride data (Fig. S2). For each stride, walking speed was determined separately. All speeds are reported in m s⁻¹ (means±s.d.).

One of the merits of XROMM is the possibility to perform ‘virtual experiments’ (Brainerd et al., 2010; Nyakatura and Fischer, 2010a; Baier and Gatesy, 2013). By measuring step length both with active and ‘muted’ (turned off) rotations of individual elements, we separately assessed the contribution of C7, scapula, humerus and ulna to step length. For each rotoscoped stride on the treadmill, we measured the distance between the position of the distal radial articular surface at TD and at LO, using the ‘distance-tool’ in Maya^®. The procedure was repeated by subsequently ‘muting’ rotations of joints individually, freezing the joints' orientation at TD. Then, the distance between the TD position and the ‘new’ position of the distal radial surface at the original frame of LO (unmuted control trials from rotoscoping) was measured. To eliminate effects of running speed or movement of the animal on the treadmill, translations of C7 (i.e. the element on top of the hierarchical chain) were muted in all assessments of the contributions of individual skeletal elements.

Elbow abduction from a dorsal view was also measured. To assess elbow abduction from the shoulder–hip parasagittal plane and compare it with previously published data (Jenkins, 1971; Lin et al., 2019a), the model was analysed from a dorsal view, as if a camera were present in that position, imitating Jenkins’ (1971) experimental setup. At TD, mid-stance and LO, a screenshot was taken and imported into ImageJ version 1.51 (Schneider et al., 2012). Elbow abduction angles were quantified using the ‘angle tool’ in ImageJ, between an artificial straight line from humeral head to humeral capitulum (COR of radius) and the shoulder–hip parasagittal plane created earlier in Maya^® using a construction plane. Note, that in this paper we refer to ‘elbow’ when describing the position of the olecranon (relative to the shoulder–hip parasagittal plane). Rotations of the ulna and radius relative to the humerus occur in the humeroulnar joint and humeroradial joint, respectively.

Kinematic accuracy and reliability are expected to vary between elements and DOFs. Lower ordered elements in the hierarchical chain (especially the ulna and radius) are subject to summation of inaccuracies on higher ordered elements, resulting in overcorrection of the lower elements (see Gatesy et al., 2010). Furthermore, highest inaccuracies are expected downwards in the rotation order of the Euler angles, as rotation of higher ordered axes affects values on lower ordered axes (see Robertson and Gordon, 2004).

To visually compare locomotion kinematics across species of different locomotor modes and body size, we used the sprawling gait space (SGS; Nyakatura et al. [2019]), which accounts for those kinematic features that have been demonstrated to contribute the most to propulsion during sprawling quadrupedal locomotion (Barclay, 1946; Edwards, 1977; Ashley-Ross, 1994; Karakasiliotis et al., 2016). It incorporates lateral spine bending, and LAR of the humerus during stance phase, and reflects body height above ground measured relative to the distance between pectoral and pelvic girdle (inter-girdle distance, IGD). Retraction of the humerus, although contributing to forward propulsion (Barclay, 1946; Edwards, 1977; Ashley-Ross, 1994; Karakasiliotis et al., 2016), is excluded from the SGS, as it has been shown to be inversely related to humeral LAR in a morphologically and phylogenetically diverse sample of sprawling tetrapod species (Nyakatura et al., 2019). Kinematic data for comparison were acquired from various sources: dog (beagle: Wachs et al., 2016; Andrada et al., 2017); two-toed sloth (Nyakatura and Fischer, 2010a); blue-tongued skink, green iguana, spectacled caiman and Mexican salamander (all Nyakatura et al., 2019); and American alligator (Baier and Gatesy, 2013). No dog spine data at the pectoral girdle were available (but see Stark et al., 2021, for a 3D musculo-skeletal model of the dog). Therefore, data from the pelvis (Wachs et al., 2016) were used instead, probably overestimating its value as the lumbar spine region usually rotates more than the pectoral girdle in mammals (see Discussion below).

Shoulder joint mobility between sprawling and parasagittal animals differs substantially both in orientation and range of motion (ROM) (Jenkins and Weijs, 1979; Jenkins and Goslow, 1983), complicating a comparison between the locomotor kinematics. Humeral LAR in sprawling animals results in the rotation of the elbow (thus facilitating retraction of the distal limb) with a fused scapula–coracoid (Fischer et al., 2010; Baier and Gatesy, 2013). Scapular and humeral LAR data of the mammals both result in internal and external elbow rotation in our and the cited model setups. Therefore, scapular and humeral LAR were summed, to allow comparison in stylopodial LAR with sprawling animals. IGDs of dog, sloth and alligator were estimated from images of a beagle (https://www.sciencephoto.com/media/1005047/view/beagle-skeleton), sloth (Nyakatura et al., 2010) and alligator skeleton (http://www.savalli.us/BIO370/Anatomy/5.AlligatorSkeleton.html), respectively, in ImageJ version 1.51 (Schneider et al., 2012). All data used in the SGS are summarised in Table S3.

The animals walked on average at 0.45±0.16 m s⁻¹ on the ground (including on the treadmill) and around 0.26±0.09 m s⁻¹ when balancing (Table 2). The hands were always placed in the shoulder–hip parasagittal plane or even further medially (closer to the sagittal plane). Hand positioning was closer to each other when the tamanduas balanced (5.9±0.16 cm versus 9.6±1.7 cm on the treadmill in the SR setup). This resulted in a combination of kinematic differences compared with the treadmill setup: the scapula was more adducted, the humerus more protracted until mid-stance, and the lower arm (radius and ulna) was constantly more adducted (see Discussion). When balancing, the hands were more pronated (Fig. 3B) and ground contact was initiated with digit III, IV and V simultaneously. On the ground, digit V initiated ground contact. During balancing, the feet were rotated more medially (Fig. 3B) than on the ground.

On the treadmill, the elbow was abducted from the shoulder–hip parasagittal plane on average 113±15, 42±6 and 30±10 deg, at TD, mid-stance and LO, respectively (Fig. 3A; N=20). Average forelimb contribution to step length by rotation of the spine at the pectoral region (here approximated by rotations of C7) was 12±7% (N=20), with almost all of this length produced by yaw and little to none by roll and pitch. The scapula contributed 51±10%, the humerus 27±7% and the ulna 6±5% to forelimb step length (N=20), all almost exclusively by retraction.

In our dataset, tamanduas displayed 72 strides of LSLC, 2 of LS single foot, 63 of LSDC and 2 of DSDC, with D ranging from 0.109 to 0.653, generally increasing on challenging supports (balancing and on inclinations) (Fig. 4A; Table S1). Speed had no effect on D (Fig. S3).

S decreased with higher walking speeds, with the lowest value of 0.54, classifying all trials as walks (Fig. S3). The negative slope of S was steeper for the pooled trials on branches (slope=−0.2484, F_1,199=63.5, P<0.001, R²=0.238) than on the ground (slope=−0.1549, F_1,161=83.7, P<0.001, R²=0.338). In the three recorded trials for −90 deg decline, the animals descended head-first, relying heavily on the prehensile tail for support. The hands were sliding down along the branch, making it impossible to discriminate between TD and LO.

DFI was linked to support type (Fig. 4B). On inclined supports and branches of 0 deg, DFI tended to be over 100; thus, hindlimb S was higher than forelimb S. DFI was on average below 100 on the ground and at −45 deg declines.

Variance of SR measurements across all analysed joints and skeletal elements was on average 4.3±2.09 deg in rotation and 0.4±0.27 cm in translation. However, there were large differences in variance between each joint, DOF and instance of stride, detailed in Table S2. While we believe the balancing data to be useful as a comparison to the treadmill data, the small sample size and high variability in the data suggest that exact values should be treated with caution. We therefore refrain from reporting exact values and compare the balancing trials with the treadmill data in general terms. An illustration of a walking cycle is presented in Fig. 5.

At the beginning of stance phase, C7 was rotated contralateral to the supporting limb by ∼10 deg yaw from the reference pose of the model setup (Fig. 6A,B). During stance, it yawed, reaching 10 deg in the ipsilateral direction at LO. Roll and pitch fluctuations were low, though C7 was pitched down constantly (∼33 deg), but less so when balancing. Dorsal displacement of the spine in the pectoral region was biphasic (Fig. S2), with highest values at mid-stance and mid-swing, averaging 16.5 cm above ground, when the shoulder passes over the manus. In balancing, the animals had lower dorsal translational values, though a pattern in the kinematic profile was less clear, with a single peak shortly after LO.

At TD, the scapula was in a maximally protracted position (−75 deg relative to the reference pose) and started to be retracted after 10% of stance phase until maximum retraction was reached at LO (−37 deg) (Fig. 6C), i.e. never reaching a vertical position of its long axis. The scapula was in the shoulder–hip parasagittal plane at TD (0 deg abducted/adducted relative to the reference pose) and was abducted towards LO where it reached its maximum abduction (at 7 deg) (Fig. 6E). The long axis was constantly rotated internally (mean −34 deg) with little amplitude of movement. When balancing, the scapula was protracted less than it was during trials on the treadmill. Furthermore, the scapula was slightly more adducted until 75% of stance phase, yet more abducted during swing phase. LAR was similar during locomotion on the two setups. The range of scapular translations in all directions was below 1 cm (Fig. S2), following a monophasic pattern.

As for the scapula, rotation with the largest amplitude was protraction/retraction (Fig. 6D). Shortly before TD (90% of swing phase), the humerus retracted until mid-stance (from 38 to ∼73 deg). Then, it protracted slightly until LO (∼68 deg), after which it retracted and reached its most retracted position (80 deg) at mid-swing. The humerus was constantly rotated internally around its long axis compared with the reference pose (Fig. 6F). Internal rotation started at 90% of swing phase, through the entire stance phase (from −15 to approximately −35 deg). The pattern in balancing was similar, yet protraction was generally lower. No data could be acquired for early swing phase in the balancing setup.

Ulna protraction displayed a biphasic curve (Fig. 7A). The ulna started in a protracted position at TD (−62 deg), after which it protracted until 25% of the stance phase (reaching −80 deg). Then, the ulna retracted until LO (to −53 deg). After LO, the ulna protracted again until mid-swing, and reached the most protracted state (at −94 deg). Abduction and LAR displayed a monophasic curve, with a peak in abduction and LAR shortly before TD (Fig. 7C). In balancing, the ulnar was generally more protracted and more adducted.

The radius displayed a similar pattern to the ulna in all rotations, but the range in abduction/adduction (30 deg compared with 20 deg in the ulna) and LAR (41 deg compared with 13 deg) was greater (Fig. 7D). When balancing, the radius was more protracted, internally rotated, and adducted (Fig. 7B).

According to Lammers and Biknevicius (2004), animals primarily adjust three factors in order to maintain stability during (arboreal) locomotion: speed, gait and limb placement. Our study confirms behavioural adjustments in all three of these aspects by T. tetradactyla when balancing on various inclined branches, compared with walking on the flat ground. The here documented reduction of speed has, for instance, similarly been reported for koalas, which are also medium-sized arboreal specialists, on more challenging supports (Gaschk et al., 2019).

Regarding utilised gaits, tamanduas almost exclusively exhibited lateral sequences (LS) (Fig. 4A). LS gaits are the most commonly observed gaits in quadrupedal mammals (Hildebrand, 1966, 1967, 1976; Cartmill et al., 2002), probably even representing the ancestral state of tetrapods (Wimberly et al., 2021). Diagonal sequences (DS) are displayed in arboreal marsupials (Schmitt and Lemelin, 2002; Shapiro and Young, 2010; Gaschk et al., 2019; Nyakatura, 2019), primates (Hildebrand, 1967; Nyakatura et al., 2008) and a few examples reported outside these groups (e.g. kinkajou; Lemelin and Cartmill, 2010). Therefore, the use of DS gaits has been regarded as an convergent arboreal adaptation to increase stability (Cartmill et al., 2002), often in combination with a more caudally oriented weight distribution (Schmitt and Lemelin, 2002). This was suggested to allow arboreal species to probe branch integrity with their forelimbs while maintaining support on their hindlimbs (Cartmill et al., 2007a).

Many primates and other fine-branch specialists increase diagonality (D) when moving from a flat surface to (simulated) branches (see Shapiro and Young, 2010, and references therein). Yet, this trend was not supported in a terrestrial marsupial (Lammers et al., 2006). Tamanduas were found here to also adjust gait when balancing on branches by increasing D. The increase of D reduces the time during which the animal's weight is supported on a unilateral bipod (thus ipsilateral hand and foot) versus the time spent on a diagonal bipod (contralateral hand and foot) and tripodal stance (Cartmill et al., 2002, 2007b), enhancing stability during balancing. We here documented higher values of D on inclining and declining supports than on a horizontal branch. In the limited comparative data available, D increased from declined to inclined simulated branches in callitrichid primates (Nyakatura et al., 2008; Nyakatura and Heymann, 2010; Hesse et al., 2015; and see discussion in Karantanis et al., 2015) and in grey short-tailed opossums (Lammers et al., 2006). For sugar gliders, higher D was reported on inclining and declining supports compared with horizontal supports, though this difference was not significant (Shapiro and Young, 2010). We propose that maintenance of high D in tamanduas for stability reasons on inclines and declines avoids the less stable LS single foot and trot (see Cartmill et al., 2002).

Whereas grasping arboreal primates and marsupials predominantly display DSDC gaits (Hildebrand, 1967; Schmitt and Lemelin, 2002; Nyakatura et al., 2008; Lemelin and Cartmill, 2010; Nyakatura, 2019) on (simulated) branches, tamanduas shift from LSLC gaits (D<0.25) on the ground to LSDC gaits (0.25<D<0.5) on branches. LSDC gait sequences on arboreal supports have also been reported for small arboreal species with less specialised grasping adaptations such as opossums, sugar gliders (Lemelin et al., 2003; Shapiro and Young, 2012; Karantanis et al., 2015), and tamarins and marmosets, which possess secondarily evolved claws (Schmitt, 2003; Hesse et al., 2015).

Furthermore, tamanduas probably additionally increase stability by use of their strong prehensile tail. Although it is not in contact with the branch during steady-state balancing, it is moved when changing direction on the narrow branches (A.S., personal observation). Probably, the positioning and movement of the tail facilitates stability by reducing whole-body momentum, as demonstrated in squirrel monkeys (Young et al., 2021) and in squirrels falling unexpectedly (Fukushima et al., 2021). While tamanduas have a larger body mass compared with the arboreal specialists displaying LSDC mentioned above, their tail probably keeps the animal's centre of mass (COM) vertically above the balancing branch and potentially shifts the COM caudally. This combination of adjustments could allow tamanduas to use the arguably less stable LSDC gaits. But, COM and ground reaction force measurements on balancing tamanduas would yield more insights.

Moreover, tamanduas display a striking shift in hand and foot posture when faced with challenging supports. When on flat ground, tamanduas characteristically walk on the lateral side of their supinated hand (Pocock, 1924; Taylor, 1978; Polania-Guzmán and Vélez-García, 2019; this study). When balancing, however, their hands are more pronated (Fig. 3B; and see kinematic results of radius, Fig. 7D), probably increasing grip on the branch. At +45 and +90 deg, the hand is placed behind the branch, ‘hooking’ with the large 3rd digit claw to allow the forelimbs to pull on steep inclines, similar to primates using their grasping hands (cf. Hesse et al., 2015). When walking on flat ground, the hindlimbs are aligned with the body, pointing forward. However, when balancing, the hind feet are rotated medially (Fig. 3B). This rotation is increased on smaller branches (P.C.D., personal observation) and probably further increases grip on the branch, aiding in keeping the COM above the branch.

Recent studies investigated the DFI and propulsive effects of forelimbs and hindlimbs on inclines. In tamarins (Hesse et al., 2015) and arboreal squirrels (Wölfer et al., 2021), increasing DFI on inclinations resulted in an emphasis of the hindlimbs' propulsive effect on inclines; and vice versa, decreasing DFI on declines emphasizes the braking role by the forelimbs on declines (Hesse et al., 2015; Wölfer et al., 2021). Similarly, increasing DFI on inclinations in tamanduas indicates a hindlimb propulsive effect on inclines and braking by forelimbs on declines.

With the animated moving models, we were able to quantify the salient abduction of the elbow in tamanduas. Compared with available data of other mammals (Jenkins, 1971; Lin et al., 2019a,b), tamanduas display the largest range in elbow abduction angles from the shoulder–hip parasagittal plane during stance phase documented so far (Fig. 3). Furthermore, the average abduction angle is larger, except compared with that of the echidna and the mole. Both are mammals displaying a sprawling gait with unique anatomical structures. For example, the monotreme echidna is characterised by a plesiomorphic shoulder girdle anatomy (Jenkins, 1970; Gambaryan and Kuznetsov, 2013; Regnault and Pierce, 2018) and the highly derived mole exhibits a permanently protracted humerus which is held in a cranially oriented position above the shoulder (Lin et al., 2019a,b). We regard this as justification for our comparison of tamanduas' locomotion kinematics with that of sprawling non-mammalian tetrapods (see below).

During stance, the spine at the pectoral girdle of tamanduas undergoes a yaw of about 21 deg. This is larger than measured in the 1st thoracic vertebra of sloths (16 deg) (Nyakatura and Fischer, 2010a) and limited otherwise available data of single vertebrae in the thoraco-lumbar region in dogs (Wachs et al., 2016), sloths (Nyakatura and Fischer, 2010b) and horses (Faber et al., 2000). Instead, tamanduas show a range of vertebrae yaw similar to that of sprawling skinks (21 deg) (Nyakatura et al., 2019) and alligators (20 deg) (Baier and Gatesy, 2013) (see Fig. 8). Correctly timed lateral bending contributes substantially to cranial displacement of an abducted elbow on the contralateral side (Baier and Gatesy, 2013; Karakasiliotis et al., 2013). Asymmetrical lateral spine bending in primates (Hildebrand, 1967) is associated with an increase in stability when applying a DSDC gait (Hildebrand, 1967; Schmidt and Fischer, 2000). With tamanduas exhibiting LS, the large symmetrical rotational range in C7 yaw tentatively suggests a link between lateral spine bending and the abducted elbow in tamanduas (see below).

In tamanduas, scapular protraction/retraction is the most pronounced rotation of the scapula, with a range of movement similar to that of sloths (Nyakatura and Fischer, 2010a) and walking dogs (Andrada et al., 2017) (Fig. 6C). Abduction of the scapula is initiated at mid-stance when the long axis of C7 (i.e. the vertebral column) is aligned with the direction of travel. This movement facilitates maintaining the shoulder in line with the direction of travel (Fig. 5A), despite subsequent lateral bending of the spine near the pectoral girdle. The shoulder is thus located more lateral to the sagittal plane at the end than at the beginning of stance phase. As a clavicle is vestigial or absent (Taylor, 1978), this positioning is likely to be a result of the scapula moving along the tapered thoracic wall, similarly described in aclaviculate raccoons (Jenkins, 1974), and sloths with a derived but present clavicle (Nyakatura and Fischer, 2010a). LAR of the scapula is almost absent; it is noticeable, however, that the scapula maintains a constant internal rotation (mean −34 deg), compared with the reference pose. This is a higher internal rotation than documented for dogs (mean approximately −10 deg) (Andrada et al., 2017) and sloths (mean approximately −22 deg) (Nyakatura and Fischer, 2010a). This positioning of the scapula thus facilitates displacement of the limbs in the medio-lateral direction during protraction/retraction.

Retraction of the humerus resembles humeral retraction in other therian mammals (Whitehead and Larson, 1994; Fischer et al., 2002; Nyakatura and Fischer, 2010a; Bonnan et al., 2016; Andrada et al., 2017). Timing during the stride is highly preserved across mammalian clades (Fischer et al., 2002) and ensures forward propulsion during the support phase and the passing of the foot above the support during the swing phase. ROM and maximal degree of humeral protraction/retraction differ substantially across clades. Cursorial mammals, such as dogs, maintain a more vertical position (20 deg humeral retraction relative to the scapula at TD) (Andrada et al., 2017). Non-cursorial mammals, such as tamanduas (40 deg), sloths (80 deg) (Nyakatura et al., 2010; Nyakatura and Fischer, 2010a) or rats (70 deg to sternum) (Bonnan et al., 2016), keep the humerus in a more retracted position, indicating a more crouched posture (Reilly et al., 2007; Biewener and Patek, 2018). With regard to humeral adduction and LAR, Jenkins (1971, 1973) investigated opossums and described adduction and internal rotation during the stance phase as a general pattern in humeral rotation of non-cursorial quadrupedal therian mammals, from which tamanduas do not appear to deviate. Humeral LAR range (∼25 deg) is similar to that of sloths (∼24 deg) (Nyakatura and Fischer, 2010a) and larger than reported for dogs (∼12 deg) (Andrada et al., 2017). Humeral internal rotation about the long axis and an obliquely shaped humeral trochlea maintain the hand in the shoulder–hip parasagittal plane during stance (Jenkins, 1973). Humeral LAR in tamanduas is probably higher as a result of the highly lateral elbow position, and higher in sloths because their suspensory locomotion places the hand medially of the shoulder–hip parasagittal plane. The configurations described above show that the pronounced abducted elbow of tamanduas is facilitated by an internally rotated scapula, retraction of the humerus and (to a lesser degree) an internally rotated humerus. This is in concordance with Taylor's (1978) observations, yet they were not quantified before.

Ulnar and radial protraction/retraction in tamanduas (Fig. 7) also follows a general pattern observed in mammals (Fischer et al., 2002). The ulna and radius show a uniform protraction/retraction (compare Fig. 7A,B), a relationship which has been reported previously in rats (Bonnan et al., 2016). Generally, the humeroulnar joint is regarded as a ‘single axis hinge joint’, confined to flexion/extension (Jenkins, 1973; Fujiwara, 2009; Bonnan et al., 2016). Yet, recent studies indicate rotation around multiple axes (Fujiwara and Hutchinson, 2012; Richards et al., 2021). For abduction, we report an amplitude of movement of up to 20 deg in the ulna and 30 deg in the radius (but note the lower signal-to-noise ratio in the lower limb data). The abduction peak is timed shortly before the medial rotation of the humerus and ensures forward positioning of the lower arm at TD. According to Taylor (1978), possibly selection occurred towards a more flexible humeroulnar joint in tamanduas, which utilise hook and pull digging (Kley and Kearney, 2007), compared with the more generalist opossum. Taylor (1978) describes the humeroulnar articulation as supported by muscular instead of bony modifications, allowing an increased ROM in ulnar abduction/adduction. Our observation is consistent with Taylor's (1978) deductions.

The radius rotates around the ulna to rotate the manus (Bonnan et al., 2016). It has been reported that tamanduas walk on the lateral side of the hand (Pocock, 1924; Taylor, 1978, 1985; Polania-Guzmán and Vélez-García, 2019). In our model setup, external LAR of the radius indicates supination of the hand (see LAR values in Fig. 7D), thus confirming previous reports for locomotion on flat ground (but see the substantial kinematic differences when balancing on a branch discussed above).

An initial model setup included the claw of the third digit. However, few data points could be acquired because of limitations of the experimental setup. Preliminary results (Fig. S2) indicate that the distal interphalangeal joint protracts and abducts the claw during stance phase. The data further suggest LAR occurs distally of the radius. Taylor (1985) explicitly describes prevention of any rotation other than protraction/retraction in the digits, to prevent dislocation. Yalden (1966) reports LAR in the midcarpal joints in Myrmecophaga but makes no comparison with tamanduas. It appears that this question remains to be thoroughly investigated, possibly with focused X-ray recordings dedicated to elucidate manus kinematics during walking and grasping.

Forward propulsion of sprawling tetrapods is mainly achieved by lateral bending of the spine and retraction and LAR of the humerus/femur (Ashley-Ross, 1994; Karakasiliotis et al., 2013; Nyakatura et al., 2019). In therian mammals, with limbs oriented in a shoulder–hip parasagittal plane (Biewener and Patek, 2018; Usherwood and Granatosky, 2020), forward propulsion is mainly achieved by caudal rotation of the most proximal pivot, thus the scapular fulcrum in forelimbs, resulting in retraction of the limb (Kuznetsov, 1985; Fischer et al., 2002; reviewed by Fischer and Blickhan, 2006). In the sprawling gait space (Fig. 8), available data for mammals (tamandua, dog and sloth) display a similar range in humeral LAR, rotating less than in the non-parasagittal animals (skink, caiman, iguana and salamander). The alligator displays an intermediate degree of humeral LAR. However, tamanduas show a higher degree in lateral spine bending than dogs and sloths, instead being in the range reported for sprawling skinks (21 deg) (Nyakatura et al., 2019) and alligators (20 deg) (Baier and Gatesy, 2013).

Maximum yaw of the spine at the pectoral girdle occurs in tamanduas shortly after TD (Fig. 6), concordant with timing of maximum yaw in alligators (Baier and Gatesy, 2013). In combination with the constant abduction of the elbow (see above), this implies timed spine bending contributes to stride extension. The manus and elbow of the limb in contact with the ground serve as pivots, propelling the contralateral forelimb forward by means of lateral spine bending, as reported for hindlimbs in salamanders (Ashley-Ross, 1994) and forelimbs in alligators (Baier and Gatesy, 2013). Given the internally rotated long axis of the scapula and humerus, this movement needs to be accompanied with ulnar/radial abduction prior to TD, to convert the C7 rotation into stride extension. This abduction is present in tamanduas (Figs 5 and 7C,D). The manus is, in contrast to ‘true sprawlers’ such as salamanders (Ashley-Ross, 1994) (Fig. 5C), always placed in the shoulder–hip parasagittal plane. The alligators' lower degree of humeral LAR in contrast to that of other sprawling animals, may be explained by the presence of a coracosternal joint, which slides caudally during stance in the sternal groove, and a mobile sternum, both contributing significantly (together ca. 20–30%) to step length (Baier and Gatesy, 2013; Baier et al., 2018).

To test the hypothesis of lateral spine bending for stride length extension in T. tetradactyla, we quantified step length contribution of the tamandua forelimb elements by rotation. To our knowledge, no comparative data on step length contribution of lateral bending in the spine region at the pectoral girdle in quadrupedal mammals is available. We report 12% for tamanduas by vertebrate yaw, which is lower than reported for non-parasagittal tetrapods [24% in agamid lizards (Peterson, 1984) and 17% in alligators (Baier and Gatesy, 2013)]. Still, it is likely to be higher than in other mammals, where predominantly sagittal spine bending (spine flexion and extension) is credited for the increase in stride length in asymmetrical gaits (Hildebrand, 1976; Fischer, 1994; Fischer et al., 2002; Schilling and Hackert, 2006; Panyutina and Makarov, 2022). However, lateral bending of the spine in mammals is not insignificant, especially during symmetrical gaits, with most studies focusing on the lumbar spine region (Jenkins and Camazine, 1977; Pridmore, 1992; Faber et al., 2000; Haussler et al., 2001; Shapiro et al., 2001; Nyakatura and Fischer, 2010b; Wachs et al., 2016). The range of lateral bending during stance in the thoraco-lumbar region of cursorial horses (<8 deg; Faber et al., 2000, 2001) and dogs (9.2 deg; Wachs et al., 2016) is lower than we report for the pectoral region in tamanduas (21 deg). This is remarkable, as lateral rotation of the spine appears to be generally larger in the thoraco-lumbar region than in the pectoral girdle in mammals (Pridmore, 1992; Fischer, 1994; Fischer and Blickhan, 2006). Sloths for example display a lateral rotational range of 28 deg in the pelvis and 16 deg in the 1st thoracic vertebra during stance (Nyakatura and Fischer, 2010a,b). In tamanduas, while rotation of the spine at the pectoral girdle contributes considerably to step length, most propulsion is generated by the scapula. Tamanduas' scapula accounts for 51% of step length, which is well within the documented range of therian mammals (43–73%; Fischer and Blickhan, 2006) and more than reported for the contribution of stylopodia (humerus/femur) in sprawling tetrapods (20–25%; Peterson, 1984; Baier and Gatesy, 2013). Tamanduas thus seem to implement the sprawling locomotor trait of spine bending into their therian kinematic pattern but otherwise maintain a typical therian forelimb kinematic profile.

This study investigated the seemingly sprawling locomotion of tamanduas. Compared with walking on the ground, tamanduas balance more slowly and cautiously, increase hand pronation to grasp the support using their large claw on the third digit, and rotate their hind feet medially on branches. Furthermore, tamanduas increase diagonality on challenging supports such as horizontal and differently inclined poles, resulting in a shift from LSLC gaits to LSDC gaits. This change probably minimises the time spent in a less stable unilateral bipod towards more time spent in the more stable bilateral bipodal and tripodal stance (Cartmill et al., 2002, 2007b). We also report an increase in DFI when balancing on poles, indicating a shift towards hindlimb-biased propulsion on arboreal supports (see Lemelin et al., 2003). In tamanduas, kinematic parameters such as protraction/retraction, abduction/adduction and LAR of the scapula and humerus are shown to be similar to those reported for other therian mammals (Fischer et al., 2002; Nyakatura and Fischer, 2010a; Bonnan et al., 2016; Andrada et al., 2017). Our results confirm the proposal by Taylor (1978) that tamanduas' characteristic abduction of the elbow results from an internally rotated scapula and a retracted humerus. Furthermore, we quantify the range in elbow abduction from the shoulder–hip parasagittal plane. While lateral bending of the spine does match the kinematics of some ‘true’ sprawling animals (Baier and Gatesy, 2013; Nyakatura et al., 2019), main forward propulsion is achieved by retraction of the scapula and the manus is placed in the shoulder–hip parasagittal plane, deviating from a typical sprawling locomotion pattern. The forelimb anatomy enables tamanduas to implement powerful lower arm retraction and twisting motions when hook and pull digging (Taylor, 1978, 1985; Kley and Kearney, 2007). The salient abducted elbow thus appears to represent primarily an adaptation to this mode of digging, allowing the application of multidirectional forces on the substrate (Taylor, 1978, 1985). Increased lateral spine bending is probably a mechanism to utilise this trait in locomotion secondarily, increasing tamandua’s step length.

The authors thank I. Schappert from the Zoo Dortmund for providing the two tamanduas and M. Roller from Tierpark Berlin for support in data acquisition. C. Funk of the Museum für Naturkunde, Berlin, provided access to tamanduas’ frontal limb bones, on which M. Müller (Freie Universität Berlin) and F. Alfieri (Humboldt Universität zu Berlin) aided in photogrammetric imaging to build 3D bone models. E. Amson (Staatliches Museum für Naturkunde, Stuttgart) provided the µCT scans of a tamandua C7, which T. Aschenbach and L. Eigen (both Humboldt Universität zu Berlin) helped to convert to a surface model. J. Wildau (Friedrich Schiller Universität Jena), J. van Beesel (Max Planck Institute für evolutionäre Anthropologie), and S. Gatesy and A. Manafzadeh (both Brown University) provided useful background regarding SR and Euler angles. S. Toussaint and L. Botton-Divet (both Humboldt Universität zu Berlin) helped in interpretation of results. L. Bormann aided during photogrammetry and designing the graphs, which D. Scheidt helped to edit. J. Wölfer (Humboldt Universität zu Berlin) aided in statistical analyses. A previous version of the manuscript was greatly improved by Armita Manafzadeh, an anonymous reviewer and the editors.