Peripersonal tracking accuracy is limited by the speed and phase of locomotion

Recent evidence suggests that perceptual and cognitive functions are codetermined by rhythmic bodily states. Prior investigations have focused on the cardiac and respiratory rhythms, both of which are also known to synchronise with locomotion – arguably our most common and natural of voluntary behaviours. Unlike the cardiorespiratory rhythms, walking is entirely under voluntary control, enabling a test of how natural and voluntary rhythmic action may affect sensory function. Here, we show that the speed and phase of human locomotion constrains sensorimotor performance. We used a continuous visuo-motor tracking task in a wireless, body-tracking virtual environment, and found that the accuracy and reaction time of continuous reaching movements were decreased at slower walking speeds, and rhythmically modulated according to the phases of the step-cycle. Decreased accuracy when walking at slow speeds suggests an advantage for interlimb coordination at normal walking speeds, in contrast to previous research on dual-task walking and reach-to-grasp movements. Phasic modulations of reach precision within the step-cycle also suggest that the upper limbs are affected by the ballistic demands of motor-preparation during natural locomotion. Together these results show that the natural phases of human locomotion impose constraints on sensory function and demonstrate the value of examining dynamic and natural behaviour in contrast to the traditional and static methods of psychological science.

shelf, intercept a passing ball, or reach to shake an outstretched hand. Previous work has 83 typically investigated reach and locomotive behaviour in isolation. When combined, a few 84 studies have focused on grasp movements when approaching a stationary object, and 85 analysed preferred foot-stance for postural stability (Bellinger et  For this study, we designed a task to assess peripersonal sensorimotor precision 100 continuously during locomotion by combining a wireless virtual reality (VR) environment and 101 the relatively new framework of continuous psychophysics (Bonnen et al., , 2017; 102 Cormack et al., 2015). While walking along a simulated path, participants made reaching 103 movements to track continuously the 3D position of an object that changed position 104 unpredictably. The tracking task allowed us to quantify changes in reaching precision 105 continuously over the step-cycle as each video frame of position-tracking provides a 106 measurement, allowing rapid collection of a dense and finely time-sampled data set. This 107 advantage allowed us to test whether tracking responses varied over time during locomotion, 108 and within the step-cycle, while manipulating walking speed. 109 110 We hypothesised that reach performance would improve at slower walking speeds 111 based on results showing that a decrease in walking speeds is observed when approaching 112 a stationary target (Rinaldi et  2014). To foreshadow our results, we found the contrary: reach accuracy was poorest when 115 walking slowly. We also found clear phasic modulations of both reach accuracy and reaction 116 time over the step-cycle. Both measures changed during the swing phase of each step-cycle 117 and in the approach to heel-strike, suggesting the continuous nature of walking imposes 118 rhythmic demands on perception-action coupling. 119 120 reach distance, target height was calibrated to 80% of participant's standing HMD elevation 169 (approximately chest height) at the start of each trial. 170 171 Procedure and Task  172 Upon arrival, participants reviewed the participant information sheet, familiarised themselves 173 with the testing environment and were given the opportunity to ask questions before 174 providing informed consent. A brief introduction to the wireless VR apparatus, hand-175 controllers and battery pack was provided, before launching a practice block to expose 176 participants to each trial combination in our design. 177 Our experiment employed a 2 x 2 factorial design, combining two walking speeds 178 (Slow, Normal) with two target speeds (Slow, Fast). On all walking trials, the walking speed 179 was set by the walking guide, which participants had to keep pace with in order to perform 180 the task. The same 9.5 metre distance was traversed at a constant velocity, with the 181 forthcoming walking pace indicated before the start of each trial. We set slow and normal 182 walking trial durations at 9 and 15 seconds, resulting in walking speeds of 0.63 m/s and 1.1 183 m/s, respectively. These speeds were set as previous research has identified that human 184 walkers in unconstrained, flat environments prefer an average walking speed of 185 approximately 1.4 m/s over long distances (Hausdorff et al., 1996)  themselves behind the red X to align themselves with the walking guide and were instructed 200 to complete the task as they had practised. The target condition was displayed in text before 201 each trial (e.g.: "On the next trial the target is SLOW, walk speed is SLOW"), and the colour 202 of the target before trial onset was changed as a visual indicator of walk speed (slow = blue, 203 normal = yellow). 204 Each walking trial began by clicking the hand-held trigger, which began the walking 205 guide's smooth linear motion at a constant velocity. During trials, the position of the target 206 moved two-dimensionally in a pseudo-random manner in the fronto-parallel plane with the 207 constraint that target position was restricted to a circle of 20 cm radius centred at the target's 208 starting location. The update frequency of the target movement was between 0.2 to 0.45 209 seconds (sampled at random from a uniform distribution with 1 ms resolution). Maximum 210 target velocity on slow target trials was set to 0.53 m/s (distance 0 to 10.6 cm), and 0.85 m/s 211 (distance 0 to 16.97 cm) on fast target trials. These parameters were chosen after careful 212 pilot experimentation to induce continuous, achievable, and yet challenging motion tracking. 213 Participants were instructed to track the target by maintaining their hand position as close to 214 the centre of the moving target as possible, while keeping pace with the walking guide. 215 Visual feedback was provided through colour changes to the target. The target was green if 216 hand position was within 6 cm of the target, and red otherwise. A video containing an 217 example trial sequence from the first-person view is located at https://osf.io/49wdt. 218 219 Data analysis 220 Each trial resulted in time-series data for head, target and hand positions. We extracted 221 step-cycle phase based on head position, and quantified task performance using the target 222 and hand positions. Most analyses were performed in MATLAB (version 2022a) using 223 custom scripts, and repeated-measures ANOVAs were performed in JASP 0.16.3.0. All raw 224 data and analysis codes are available on the repository https://osf.io/jdpwc/. 225 226 Gait extraction from head-position data 227 Walking results in a highly regular oscillatory pattern of movement (Hausdorff et al., 1996), the time-series. We excluded from all analyses any data occurring during the first two and 236 last two steps of each trial, to allow for changes in acceleration at the endpoints of our walk 237 trajectories. Gait percentages are typically measured from heel-strike to heel-strike of the 238 opposite foot (step-cycle), or heel-strike to heel-strike of the same foot (stride-cycle). In our 239 case, as we have epoched based on the loading phase of each step, heel-strike occurs 240 slightly before the trough in the envelope of the vertical centre-of-mass (Gard et al., 2004;241 Hirasaki et al., 1999). In the present work, we have labelled step-cycle completion from 1-242 100% based on these loading-phases in our epoched time-series. Approximate locations of 243 heel-strike and toe-off are also indicated based on prior studies which have measured foot 244 pressure and head acceleration simultaneously (Mulavara & Bloomberg, 2002), at 10% step-245 cycle timing before and after mid-stance, respectively (MacNeilage & Glasauer, 2017). 246 We additionally performed a series of pre-processing steps which visually identified 247 individual trials for exclusion. Each individual trial was visually inspected for data 248 discontinuities (a result of wireless transmission delay or slip), transient spikes in error (a 249 result of disengaging from the task), and poor gait-extraction (a result of vigorous head 250 movement). Over all participants, an average of 7.1 trials (SD = 8.78) were rejected in this 251 manner. 252 For our step-cycle based analyses, we additionally identified foot placement based 253 on the direction of lateral sway (e.g., left foot stance and right foot swinging when head 254 position is tilted to the left of centre). We then resampled the raw time-series for head, hand 255 and target position to 100 data points (1-100% step-cycle completion). This resampling 256 procedure is common in gait and posture research to align step and stride-cycle epochs 257 when walking at a steady speed (MacNeilage & Glasauer, 2017), and enabled us to assess 258 performance relative to position in the step-cycle. We note that our main results hold for the 259 raw time-series data without resampling, owing to the highly regular step length and step 260 duration entrained during steady-state locomotion. Figure 4A-C shows key points in this 261 workflow. 262 263

Error based on hand-target Euclidean distance 264
As the actual target position was known on every frame, we calculated error based on the 265 Euclidean distance between actual target position and current hand position per time-point. 266 For our condition comparisons, we quantified overall performance using the Root Mean 267 Square Error (RMSE) of these distances. RMSE is common in regression analyses, and 268 measures the average distance between the predicted values and the actual observations 269 from the line of best fit. In our case, the observations about an idealised line of best fit 270 correspond to the hand position and target position, respectively, and RMSE enables a 271 quantification of error between these time-series. For our condition comparisons (Figure 2), 272 RMSE was first calculated per trial, and then averaged within conditions, per participant, 273 before performing within-participant comparisons at the group level. 274 275 Error relative to target location 276 To quantify spatial changes in error, we analysed the distribution of recorded target locations 277 on the frontal dimension (parallel to the participant's coronal plane), as well as average 278 RMSE per position. This analysis first tallied the distance between hand and target position 279 at each location, before calculating RMSE based on the number of frames (i.e., duration) 280 that the target spent at each location. For this analysis, we resampled the target-location 281 space into a 61 x 61 grid (1 cm resolution), centred at the starting target location of each 282 trial. At each location we quantified RMSE per trial, and then averaged within trial types, and 283 across participants. For visualisation and analysis purposes, group-level effects were 284 restricted to locations which contained data from all participants. 285 To compare the magnitude of error across target locations, we performed a non-286 parametric shuffling analysis to create a null distribution that removed the consistency of 287 target location across participants. This was based on similar analyses used to compare the 288 spatial/topographic distribution of MEG/EEG activity (Maris & Oostenveld, 2007). This 289 analysis tests whether the observed group-average error at a specific location is greater than 290 can be expected by chance. On 1000 permutations per participant, we changed the row and 291 column index of each observed RMSE location, selecting a new row-column combination 292 from two uniform distributions (with replacement). These distributions contained all possible 293 row and column locations per participant, removing the correlation between target-location 294 and error magnitude on each permutation. Across participants, we then compared our 295 observed data to the 95% Confidence Interval (CI) of average RMSE from shuffled locations, 296 and determined whether observed error was greater than expected by chance when falling 297 beyond the bounds of this null distribution. The results of this analysis, and bounds of the 298 null distribution are shown in Figure 3. 299 300 Error relative to the phase of locomotion 301 For our step-cycle based analysis, we first quantified RMSE at all time-points within a trial, at 302 a fixed position in the step-cycle (from 1-100%). This analysis produced an average RMSE 303 per trial over the step-cycle, and enabled us to compare relative RMSE at different phases of 304 locomotion, both within and across conditions. 305 Of central interest was whether, and when, differences in continuous reach-error 306 would emerge over the step-cycle when comparing between target speeds and walking 307 speed conditions. To statistically compare step-cycle based RMSE between walking speeds, 308 we performed a series of paired-samples t-tests at each location in the step-cycle. We 309 visualise significant differences between fast and slow walking speeds by highlighting step-310 cycle positions with p < .05 after false discovery rate (FDR) corrections (

Cross-correlogram (CCG) and windowed Cross-correlogram (wCCG) 314
As a complement to distance based error, we also computed cross-correlograms (CCGs) as 315 a proxy for reaction-time, based on the correlation between hand and target time-series. 316 CCG functions plot the correlation between the time-series of a target and response as a 317 function of the temporal-lag between them (e.g., Bonnen et al., 2015Bonnen et al., , 2017Mulligan et al., 318 2013). To compute the CCG, we converted the position data to a velocity time-series for the 319 hand and target on both axes (vertical and horizontal). We computed single-trial CCG 320 functions after omitting the first and last steps (as described above), before averaging within 321 conditions and across participants. We retained the time-lag in the peak of each CCG 322 function as our proxy for reaction-time per condition. 323 In addition to single-trial CCGs, we computed short-interval windowed CCGs (wCCG) 324 to assess reaction-time relative to position in the step-cycle. The wCCG method analyses 325 the cross-correlation function over a short sliding window, and was originally developed to 326 account for dependencies between behavioural time-series which may not be stable over 327 time (Boker et al., 2002;Roume et al., 2018). This analysis followed a three step process. 328 We computed wCCG with a short sliding window using the corrgram function (Norbert, 2007) 329 (10 samples, approximately 110 ms duration, 1 sample overlap), and retained the wCCG 330 function for each time-step in each trial (omitting early and late steps as described above). 331 As a second step, each wCCG was matched to a simultaneous step within each trial, and 332 based on the wCCG time-step was allocated to a bin representing location within the 333 simultaneous step-cycle. For analysis, we allocated wCCG functions into quintiles, when 334 falling within the 1-19%, 20-39%, 40-59% 60-79%, and 80-100% quantiles of a single step. 335 As a final step, we averaged the wCCG functions within each step quintile, and compared 336 their relative lag in peak correlation amplitude as a proxy for changes in reaction-time over 337 the step-cycle. 338 339 Data Visualisation 340 We have implemented the raincloud toolbox (

348
We developed a continuous motion tracking task that varied walking speed 349 (slow/normal) and target speed (slow/fast) in a 2x2 factorial design. Participants were 350 instructed to minimise the distance between their dominant right hand and a moving target, 351 while steadily walking within an enclosed 9.5 m wireless virtual reality environment ( Figure  352 1). The moving target followed a pseudo-random path within a comfortable reaching 353 distance, allowing us to assess reach accuracy in peripersonal space, and changes in task 354 performance over the step-cycle. Decreased reach error at normal walking speed 368 369 As expected, our instruction to walk at a slower pace resulted in participants adopting 370 a conservative gait, with changes evident in both step duration and step distance (Figure 2).

371
A series of 2 (walking speed) x 2 (target speed) repeated-measures ANOVAs revealed that 372 when walking slowly, step duration significantly increased (F(1,24)  hoc tests examining the effect of target speed indicated that the fast target speed led to 377 decreases in both step distance and duration and thus a more conservative gait. 378 Having established large differences in gait parameters while walking and tracking 379 targets at different speeds, we next quantified reach error based on the Euclidean distance 380 between the time-series of hand and target positions (see Methods). As prior research has 381 shown that walking speed often slows to enable prehension movements (i.e., reach-to-382 grasp), and that walking slowly can offset cognitive demands in dual-task scenarios, we 383 hypothesised that our continuous reach-tracking task would be performed better at slower 384 walking speeds. Error results shown in Figure 2G, however, indicate that overall, and in 385 contrast to our hypotheses, reach error increased when walking slowly. As expected, when 386 tracking the fast target, reach error also increased. A repeated-measures ANOVA revealed 387 significant main effects of walking speed (F(1,24) = 8.66, p = .007, ηp 2 = 0.27) and target 388 speed (F(1,24) = 226.26, p < .001, ηp 2 = 0.90) on reach error, with no interaction (p > 0.2). 389 We next performed an additional analysis to determine whether reach error was 390 influenced by left or right foot support (Figure 2H), as prior research has indicated a task-391 dependent preference for ipsilateral or contralateral foot placement when reaching for a 392 stationary object. When extending our repeated measures ANOVA to include support foot (2 393 x 2 x 2; walk speed x target speed x support foot) a significant three-way interaction was 394 found (F(1,24)

Reach error increases in the contralateral peripersonal space 420 421
We next investigated reach error relative to the target's location in peripersonal space. 422 Analysing across all trials, we observed that targets were more likely to occur centrally than 423 peripherally ( Figure 3A). This central tendency was expected because the random-walk 424 algorithm reset after every trial so that all targets began at the centre of the motion 425 boundary. Figure 3A also illustrates a clear symmetry of target locations around the origin 426 which contrasts markedly with average target error per location (Figure 3D), which 427 increased when reaching laterally across the body into the contralateral peripersonal space. 428 Reaching equivalent distances in the vertical dimension did not lead to an analogous 429 increase in error. 430 To quantify these changes in error, we calculated the average error on either the 431 horizontal or vertical dimension, and compared the observed error at each horizontal or 432 vertical position with a null-distribution resampled from all locations ( Figure 3D, see 433 Methods). This analysis revealed that performance was significantly worse when reaching 434 into contralateral peripersonal space (p < .001), and significantly improved on the ipsilateral 435 side (p < .001), compared to the 95% CI of error sampled from all locations. There were no 436 significant changes in error when reaching along the vertical dimension. 437 Given the asymmetry in reaching errors was in the horizontal axis, we analysed 438 whether this difficulty reaching accurately into contralateral space was mediated by whether 439 participants were supported by their left or right foot. We hypothesised that if the 440 contralateral error was driven by kinematic or postural demands, then this difference may be 441 larger when swinging the foot ipsilateral to the reaching hand (cf. Figure 2H). Contrary to 442 this expectation, there was no difference in error across peripersonal space based on stance 443 (Supplementary Figure 1). Upon visual inspection, the effect on the contralateral side was 444 not present in each walking speed x target speed condition and was strongest in the most 445 challenging condition: slow walk with fast target (sWfT: Figure 3E). The data in this condition 446 show that walking slowly increases the spatial distribution of reaching error overall and 447 especially when reaching into contralateral peripersonal space. The reduced ability to reach 448 contralaterally arises despite the greater postural stability afforded by a slower more 449 conservative gait, a result we return to in our Discussion.  We have shown that overall reach error is greatest when walking slowly, and when 476 reaching to the contralateral side. We next delved further into this overall error by performing 477 an analysis to quantify whether error changed over the step-cycle. For this, we used 478 Euclidean error (jointly determined by the vertical, horizontal and depth dimension) as a 479 measure of tracking performance and all conditions were examined. We also compared the 480 absolute reach error on each movement dimension, and time-course of reach error when 481 supported by the left or right foot.. 482 The time-course of Euclidean reach error is shown in Figure 4H. While the overall 483 pattern of Figure 2 is preserved (i.e., higher error when walking slowly, tracking fast targets), 484 the difference in reach error when walking at slow vs fast speeds is shown to oscillate, 485 reaching a maximum at the beginning of the swing-phase of each step. We statistically 486 evaluated these error differences with a series of paired-sample t-tests at each time-point in 487 the step-cycle. For slow-target conditions, the difference in error when walking at different 488 speeds began immediately prior to footfall, and was protracted through the first half of each 489 step (clusters 1st -45th, 79th -100th percentiles, pFDR < .05). For fast-target conditions, 490 differences in error based on walking speed were confined to a shorter period (cluster 14th -491 32nd percentile, pFDR < .05). 492 As a complement to the Euclidean error, we also quantified fluctuations in error 493 magnitude upon the vertical, horizontal, and depth axes separately, finding distinct patterns 494 in each dimension across the phase of locomotion. Vertical error was modulated roughly 495 symmetrically over the step-cycle, peaking around the midpoint of each step before falling to 496 a baseline level around the time of footfall. In contrast, horizontal error was sinusoidally 497 modulated, and peaked in the ascending phase of the step-cycle, regardless of left-or right-498 foot support. Error in the depth dimension was relatively constant, presumably due to the 499 smooth and predictable linear motion of our walking guide. Figure 4D displays the time-500 course of reaching error over the step-cycle for each of these conditions. 501 Together, these results demonstrate that condition-average differences in target 502 tracking based on target walking speed are not stationary, with phasic modulations and 503 optimal periods of sensorimotor precision over the step-cycle. 504 505 506 respectively. E) Euclidean error over the step-cycle. Red shading represents significant 519 differences between walking speeds, within a target speed condition. Large differences (not 520 visualised) remained between target speed conditions (p < .05, FDR corrected). 521 522 Cross-correlogram analysis: Walking slowly increases reaction times 523 524 We performed a series of cross-correlogram (CCG) analyses to serve as a proxy for 525 reaction time over the step-cycle. Cross-correlograms estimate the correlation between two 526 time-series over multiple time-points, by shifting the time lag (temporal offset) between them. time lag at which the correlation peaked in each condition (Figure 5). 530 We measured the CCG lag separately on the horizontal and vertical axes, and 531 observed a pattern of reaction times that indicated a speed-accuracy trade-off for fast-532 targets. Specifically, although distance-based error was larger in these conditions (cf. Figure  533 2 and 4) reaction times were also faster than when tracking the slower moving targets 534 (Figure 5). Participants were also faster overall when walking at normal speeds. We 535 statistically evaluated the difference in peak CCG lag between conditions in a 2x2x2 ANOVA 536 (walk speed x target speed x axis). Significant main effects were observed for walking 537 speed (F(1,24) = 9.51, p = .005, ηp 2 = .29), and target speed (F(1,24) = 40.87, p < .001, ηp 2 538 = .63). We also observed a main effect of motion axis (F(1,24) = 49.18, p < .001, ηp 2 = .67), 539 which indicated that responses were slower overall when correcting for position shifts on the 540 vertical axis. There was additionally an interaction between target speed and motion axis 541 (F(1,24) = 8.49, p=.001,ηp 2 = .26), such that the difference in reaction times between target 542 speeds was greater for vertical movements. After observing that accuracy fluctuated over the step-cycle (Figure 4), we next 561 turned to whether reaction-times also changed based on the relative phase of locomotion. 562 Accordingly, we applied a windowed cross-correlogram analysis (wCCG), to assess whether 563 the lag in peak CCG function varied over the step-cycle. The wCCG method focuses on a 564 short sliding window, and was originally developed to account for dependencies between 565 behavioural time-series which may not be stable over time (Boker et al., 2002;Roume et al., 566 2018). Here, we calculated wCCG functions between the time-series for hand and target 567 velocities, for both the vertical and horizontal axes, and assigned wCCGs to a relative 568 position in the step using a three step process (see Methods for details). In brief, we first 569 computed the wCCG within a short, sliding window of 111 ms, with 11 ms overlap. As a 570 second step, each wCCG was matched to its simultaneously occurring step, and the wCCG 571 function was allocated to a bin representing relative position in that step-cycle. Each wCCG 572 functions was allocated to a step quintile (i.e., for when the centre of each sliding window fell 573 within the 1-19%, 20-39%, 40-59 % 60-79%, and 80-99% quantiles of a single step, 574 respectively) and averaged within each quintile. The comparison of their time-lag to wCCG 575 peak then served as a proxy for changes in reaction-time over the step-cycle. 576 Our analyses revealed significant changes in wCCG lag over the step-cycle ( Figure  577  6). There was a significant interaction between motion axis and step quintile (F(4,96) = 6.32, 578 p 3.39, p = .012, ηp 2 = .12), post-hoc analyses revealed that reaction times were fastest at the 588 beginning of the swing phase of each step, before increasing throughout the step-cycle until 589 the approximate heel strike (q1 vs q4, t(24) = -3.46, p = .020). 590 In sum, these analyses show that the spatio-temporal tracking speed of a target is 591 fastest during the swing phase of each step, before slowing around the preparation time for 592 the next step, prior to heel strike. 593 594 595 596 We combined motion-tracking, continuous psychophysics and wireless virtual reality 609 to investigate how steady-state human locomotion affects sensorimotor performance. We 610 operationalised sensorimotor performance by computing accuracy and reaction time 611 measures from a continuous reaching task, and found that normal walking speeds were 612 advantageous. Reach error and reaction times both improved when walking at a normal 613 pace, in contrast to our expectations based on prior research. We also observed phasic 614 modulations to both reach accuracy and reaction time during locomotion. We interpret these 615 results with reference to recent work highlighting that the step-cycle imposes rhythmic 616 changes to sensory processing, potentially due to the ballistic demands of motor-preparation 617 and been shown between left and right hands when executing simultaneous reaching tasks 661 (Kelso et al., 1979). 662 Whether limited by biomechanics or cognitive demands, our results suggest that 663 walking slowly is detrimental to reach performance in a continuous tracking task. In addition 664 to these condition based differences, distinct alterations to reach performance also occurred 665 over the step-cycle, at both walking speeds. 666 667 Optimal accuracy and reaction times during distinct phases of locomotion 668 669 As a complement to the speed-dependent change in reach accuracy outlined above, 670 we also observed that reach error was dictated by the phase of the step-cycle, during both 671 normal and slow walking conditions. This is in contrast to some previous work, which has 672 described how relative to an end-point in world-space, the trajectory of reaching behaviour is 673 identical whether standing or walking (Marteniuk et al., 2000;Marteniuk & Bertram, 2001). 674 This 'motor equivalence' was taken as evidence that reach behaviour was capable of 675 perfectly compensating for the postural changes introduced by walking, as a control strategy 676 to cope with reaching in dynamic contexts. In our data, reach performance was relative to an 677 end-goal in world space (the moving target), yet clear oscillations in error occurred during 678 the step-cycle, in contrast to the motor equivalence hypothesis (Figure 4). More specifically, 679 reach error was largest during the first-half of each step, in the approximate swing-phase, 680 and decreased prior to the time of heel-strike. These phasic changes were present in each 681 walking speed and target speed combination. We also observed that reaction times, 682 computed from the lag-to-peak in windowed cross-correlogram functions, were fastest for 683 movements early in the step-cycle ( Figure 6). We note that while small, the maximum 684 difference in reaction times over the step-cycle was 22 ms (vertical motion, quintile 2 vs 5), a 685 similar magnitude to the facilitatory effects of spatial cueing on visual detection (Chica et al.,686 2014). 687 Our evidence of phasic changes to reach performance over the step-cycle extends 688 prior work investigating the initiation of discrete upper-limb movements during locomotion. 689 Carnahan et al. (1996) recorded the envelope of upper-limb electromyographic (EMG) 690 activity when walking alone, and compared this activity to combined walking with reaching, 691 and standing and reaching conditions. Arm movements were fastest when participants were 692 walking, and reach movements during walking were superimposed upon the profile of EMG 693 activity recorded during walking only conditions. This overlap was taken as evidence that 694 reach behaviour was entrained to the normal rhythm of the step-cycle, to capitalise on the 695 oscillatory movement of upper limbs during normal walking. Earlier work has also found that 696 voluntary and discrete upper limb actions are preferentially initiated at specific phases of the 697 step-cycle. Nashner and Forssberg (1986)