Optimization of energy and time predicts dynamic speeds for human walking
Abstract
Humans make a number of choices when they walk, such as how fast and for how long. The preferred steady walking speed seems chosen to minimize energy expenditure per distance traveled. But the speed of actual walking bouts is not only steady, but rather a timevarying trajectory, which can also be modulated by task urgency or an individual’s movement vigor. Here we show that speed trajectories and durations of human walking bouts are explained better by an objective to minimize Energy and Time, meaning the total work or energy to reach destination, plus a cost proportional to bout duration. Applied to a computational model of walking dynamics, this objective predicts dynamic speed vs. time trajectories with inverted U shapes. Model and human experiment (N=10) show that shorter bouts are unsteady and dominated by the time and effort of accelerating, and longer ones are steadier and faster and dominated by steadystate time and effort. Individualdependent vigor may be characterized by the energy one is willing to spend to save a unit of time, which explains why some may walk faster than others, but everyone may have similarshaped trajectories due to similar walking dynamics. Tradeoffs between energy and time costs can predict transient, steady, and vigorrelated aspects of walking.
Editor's evaluation
This valuable study presents a new optimal control cost framework to predict features of walking bouts, adding a cost function term proportional to the duration of the walking bout in addition to the conventional energetic expenditure term. The authors show solid evidence that predicted optimal trajectories from simulations qualitatively match walking data from human subjects. This study provides a foundation for future studies testing the validity of the energytime hypothesis for human gait speed selection across different walking bouts in ablebodied and patient populations.
https://doi.org/10.7554/eLife.81939.sa0Introduction
Many aspects of human walking are determined by minimization of metabolic energy expenditure. For example, the preferred step length (Atzler and Herbst, 1927) and step width (Donelan et al., 2001) minimize energy expenditure for a given steady speed, and the preferred steady speed approximately coincides with minimum energy expenditure per distance traveled (Figure 1A, Ralston, 1958). This speed, as well as the economy of walking, both decline with age, disability, or poor health. As such, preferred speed is widely employed as a clinically useful indicator of overall mobility (Afilalo et al., 2010; Studenski et al., 2011). However, there are naturally many other factors that also influence walking. All walking tasks have a beginning and end, and some may spend little or no time at steady speed. Some tasks may also occur with a degree of urgency, and some individuals may habitually walk faster than others, for reasons not obviously explained by economy. Energy economy is a powerful and objective explanation for steady walking speed, but it does not readily accommodate these everyday observations. Realistic walking tasks must therefore be governed by more than energy economy alone.
The specific energy measure thought to govern steady walking speed is the gross metabolic cost of transport (COT). Defined as energy expended per distance travelled and body weight (or mass), it has a convex dependency on speed. Its minimum (termed minCOT here) seems to predict the steady preferred speed of about 1.25 ms^{1}, as widely reported for laboratory settings (Ralston, 1958; Martin et al., 1992; Willis et al., 2005; Browning and Kram, 2005; Browning et al., 2006; Rose et al., 2006; Entin et al., 2010), and observed of some other animals (Hoyt and Taylor, 1981; Langman et al., 2012; Watson et al., 2011). However, much of actual daily living also involves relatively short bouts of walking (Figure 1B), with about half of daily bouts taking less than 16 steps as reported by Orendurff et al., 2008. Such bouts, say of distance $D$ (Figure 1C), may spend substantial time and energy on starting from and stopping at rest, and relatively little time at steady speed. For example, in short bouts of walking up to about a dozen steps, peak speed is slower than the steady optimum, and only attains that value with more steps (Seethapathi and Srinivasan, 2015). There is a substantial energetic cost to changing speeds that could account for 4–8% of daily walking energy budget (Seethapathi and Srinivasan, 2015). If energy economy is important for walking, it should apply to an entire walking bout or task, and not only to steady speed.
Another important factor for walking is time. Time is valuable in part because energy is always expended even when one is at rest (Jetté et al., 1990), and because walking faster can save time to reach destination, but at greater energy cost (Ralston, 1958). Time is also subjectively valuable, because the urgency of a task, or even of an individual’s personality, surroundings, or culture, could influence their speed. It has long been observed that people walk faster in big cities than in small towns, by a factor of more than twofold (about 0.75–1.75 ms^{1}), or about ±40% of 1.25 ms^{1} (Bornstein and Bornstein, 1976). Perhaps population density affects a person’s valuation of time (Bornstein, 1979; Levine and Bartlett, 1984; Li and Cao, 2019). Time is certainly a factor in deciding whether to walk or run (Summerside et al., 2018), and is considered an important factor in the general vigor of movements, beyond walking alone (Labaune et al., 2020). It is clearly worthwhile to expend more energy if time is of essence.
There are, however, challenges to incorporating time into walking. One method is to factor time into the equivalent of temporally discounted reward (Shadmehr et al., 2010), which refers to offering a reduced reward for longer durations, typically employed in fields such as movement vigor, foraging theory (Green and Myerson, 1996), and reinforcement learning (Sutton and Barto, 2018). Another is to express time as a cost that increases for longer movement durations, trading off against greater energy cost for shorter durations. Both the energy cost for an entire walking bout, plus a cost for time duration, could thus be combined into a single objective function to be minimized (Wong et al., 2021). This presents a second challenge, which is how to determine the optimum. Unlike the case of steady walking at a single speed (Figure 1A), an entire walking bout requires a timevarying trajectory of walking speed. This cannot be determined from the cost of transport curve, but can potentially be predicted by a quantitative, dynamical model. Simple models of walking (Figure 7), based on the pendulumlike dynamics of walking, can predict aspects such as optimal step length and step width (Kuo et al., 2005) for a steady speed, and optimal speed fluctuations for uneven steps (Darici and Kuo, 2022a; Darici and Kuo, 2022b). It remains to be determined whether they can predict the energetics and timing of walking bouts with transient conditions.
The purpose of the present study was to test whether the combined costs of energy and time can predict dynamic variation in walking speed. We propose a basic quantitative objective function called the EnergyTime hypothesis, which includes a cost for total energy expenditure or mechanical work for a walking bout, plus a penalty increasing with the bout’s time duration. We apply this objective to a computational walking model, using dynamic optimization to predict dynamic speed profiles for walking bouts of varying distance (Figure 1D). For relatively short walking bouts, this hypothesis predicts speeds that vary within a bout, and speed profiles that vary across bout distances. For longer distances, it predicts a steady walking speed, not as an explicit outcome but rather as an emergent behavior. To test these predictions, we performed a human subjects experiment, comparing empirical speed profiles against model predictions. If the model is able to predict human speed profiles, it may suggest that a valuation of time and energy can influence walking, and thus be compatible with walking bouts of any distance and any degree of urgency.
Results
Model predictions
A simple model of walking dynamics predicts theoretically optimal speed trajectories and walking bout durations. The EnergyTime hypothesis is that humans perform walking bouts that minimize an objective including the total energy and time expended for the bout. The dynamic optimization problem may be summarized as
where the total metabolic energy expenditure is evaluated for the entire walking task, and the time duration is weighted by a metabolic energy coefficient ${C}_{T}$ (in units of energy per time). In the model, positive mechanical work is used as a proportional indicator of human energy expenditure, with (lowercase) work coefficient ${c}_{T}$. This coefficient is a valuation of time, and may be interpreted as the energy or work one is willing to expend to save a unit of time. The overall objective is to be minimized with an appropriate trajectory of the model’s speed, which in the human is the outcome of active control actions. The optimal control actions are subject to constraints, namely the specified distance of a walking bout and the governing walking dynamics (see Methods for details). Walking dynamics refers to the dynamics of the body, where the stance leg behaves like an inverted pendulum and the swing leg like a swinging pendulum. These dynamics also describe the mechanical work and energy associated with a speed trajectory, and how long each step takes. Step length was nominally kept fixed, and then varied in parameter sensitivity studies below. The time duration $T$ of a bout is the outcome of the optimization, where greater valuation of time ${C}_{T}$ favors shorter duration.
The optimization predicts the speed profiles for a representative, tenstep task (Figure 2A). To focus on Energy first, the duration is kept fixed here. The EnergyTime objective predicts a gradual increase in speed, with a gently rounded profile that peaks midway through the bout. For this relatively short distance, little or no time is spent at steady speed. This contrasts with two other possibilities, to maintain steady speed at minCOT (Ralston, 1958), or to maintain steady acceleration. The steady minCOT objective produces a speed profile resembling a trapezoid, accelerating immediately to attain a fixed steady speed, maintained throughout the bout, before terminating just as quickly. Steady acceleration causes speed to increase linearly over time until peaking midbout, followed by a linear decrease back to rest. Here, all three alternatives are directed to walk the same distance in the same time, but at different costs.
Examination of the positive work inputs reveals why EnergyTime is least costly (Figure 2B). Its gentle acceleration requires moderate pushoffs, which trail off over time as the model nearly coasts to a stop at destination, taking advantage of each step’s collision loss to reduce speed at little cost. In contrast, the steady minCOT objective pays a high cost to initiate gait, and then a moderate and constant amount of work for all pushoffs. It also does not take advantage of coasting to a stop. Steady acceleration pays a high cost to peak at a high speed, which is not made up for by greatly reduced pushoffs as it comes to a stop. Some intuition may be gained by considering the analogous situation of a vehicle driving a short fixed distance between two stop signs, in fixed time. It is generally economical to accelerate and decelerate gradually, and not necessarily maintain steady speed except beyond a certain distance. A trapezoidal (minCOT) speed profile is not economical, because considerable energy is spent in fast acceleration, and braking maximally at the end is more wasteful than lifting off the accelerator early and coasting. A triangular (steady acceleration) profile is also not favorable, due to the work needed to briefly attain a high speed. Of course, walking and driving have different dynamics, but both have similar energetic loss rates that increase approximately with the cube of speed. The higher losses incurred at greater speeds is an important reason for the Energytime optimality of a rounded speed profile. For this task, the minCOT hypothesis ultimately costs 11% more total work, and the steady acceleration hypothesis 31% more, than minimizing EnergyTime.
Having established the energetic advantages of the EnergyTime hypothesis, we next examine how the optimal speed profiles vary with Time and other model parameters (Figure 3). Here, there are three parameters of interest: the value of time ${c}_{T}$, step length, and walking bout distance. We considered step lengths $s$ fixed at nominal (0.68 m), at slightly shorter and longer lengths (0.59 m and 0.78 m), and increasing with speed according to the human preferred step length relationship (see Methods for details). We also considered bouts of one to twenty steps, or about 0.68 m–13.7 m, as well as time valuations ${c}_{T}$ ranging tenfold, 0.006–0.06 (dimensionless). Regardless of the combination of these parameters, a few characteristics emerge. The speed profiles generally retain a gently rounded profile (Figure 3A), smoothly accelerating from rest and leveling off at a peak speed before decelerating back to rest. Unlike the trapezoidal, minCOT prediction, the human speed profiles are always peaked, particularly for short bouts. The longer the distance, the greater the peak speed (unlike minCOT), and the more sustained that peak, contrast to more rounded speed profiles for shorter distances. The acceleration and deceleration slopes increase slightly with longer bouts, and only for distances of about 10 m or more is there a steady gait near peak speed. The peak speed also initially increases sharply with walking distance (Figure 3B), but then approaches an asymptote for greater distances, as the cost of acceleration and deceleration becomes inconsequential to overall cost $J$. In fact, the asymptotic peak speed for long walks is a steady speed, not unlike the minimumCOT speed. But for finite walk distances, the speed profile generally does not agree with the steady minCOT hypothesis, because it varies with bout distance, and dynamically within each bout.
Another feature of the EnergyTime prediction is consistency with respect to parameter values (Figure 3A). The main free parameter is the time valuation ${c}_{T}$, for which higher values call for higher peak speeds, and therefore shorter walking durations. But with peak speeds ranging more than twofold (Figure 3A, inset), the speed profiles all had similar shape. In fact, scaling each of the profiles in time and amplitude yielded a very similar family of trajectories regardless of parameter values (Figure 3A). This is also the case for variation in step length, with nominal, long, and preferred human step lengths all producing similar trajectories. Similarly, the peak speed vs. distance curves resembled a saturating exponential regardless of parameter values (Figure 3B), and these were also scalable in amplitude to yield a single family of curves. Walking durations vs. distance (Figure 3C), also had similar, scalable and curvilinear shape for all parameters. Similar profiles are produced regardless of whether the model takes step lengths that are fixed, or that scale according to the empirical step length vs. walking speed relationship for steady walking (insets, Figure 3A). We therefore subsequently keep step length fixed (equivalent of 0.68 m for human) for simplicity. As a result, the time cost coefficient ${c}_{T}$ is effectively the model’s sole free parameter, and the predicted speed profile shapes scale very consistently with respect to that parameter.
There are thus three main predictions from the model that can be tested in human. First, the speed profiles should fall within a single consistent family, which includes more rounded shapes for short walks, and flatter for longer walks (Figure 3A). These profiles should exhibit selfsimilarity, and be scalable in peak speed and time to resemble a single, relatively uniform family of profiles. Second, the peak speed should increase with distance, with an approximately exponential saturation toward an asymptote (Figure 3B). Again, that relationship is expected to be scalable by peak speed, and testable by a single saturating exponential. And third, walking durations should increase with distance, in a slightly curvilinear relationship (Figure 3C) approaching a straightline asymptote for longer distances. For shorter distances, much of the time should be spent accelerating and decelerating, compared to relatively brief cruising periods that become proportionately greater with distance. We thus treat the time valuation ${c}_{T}$ as an empirical parameter that mainly affects the scale, but not the shape of the speed profiles and dependency on distance.
Experimental results
The human speed profiles for all trials and all distances were found to exhibit consistent profiles between subjects and between individual trials (Figure 4). These profiles resembled predictions from the EnergyTime hypothesis. Qualitatively, humans produced inverted U profiles similar to model, with sharper and lower peak speeds for shorter bouts. Longer bouts had higher and flatter peaks, where a steady speed could be discerned. Each individual subject walked at a somewhat different speed and for a somewhat different time (Figure 4A). For example, the range of peak speeds across subjects, observed for the longest (12.7 m) bout, was 1.21 ms^{–1}.78 ms_{–1}, and the corresponding range of durations was 8.51–11.86 s. Nevertheless, the profile shapes were all quite similar across subjects. In contrast to the minCOT hypothesis, the human peak speeds increased with distance, many well below the minCOT speed of about 1.25 ms^{–1}. In addition, the human speed trajectories did not resemble the trapezoidal profiles of the steady minCOT hypothesis for all distances, nor the triangular profiles of steady acceleration.
The experimental speed trajectories were scalable in speed and time, to yield a selfsimilar family of trajectories (Figure 4B). Each individual’s trajectories were normalized in time by the duration for that subject’s longest bout, and in speed by the maximum speed of their longest bout. These were then rescaled to match the average duration and peak seed across subjects, to yield a normalized set of speed profiles for all subjects (Figure 4B). The resulting normalized trajectories reveals considerable similarity between individuals, with a single, relatively uniform family of profiles for all subjects. Thus, the peak speed and duration of a walking bout of 2 m was consistently related to one of 12 m, and vice versa.
This scalability may be quantified in terms of peak speeds and durations. Examining the peak speed for each distance reveals a consistent pattern (Figure 5A). Peak speeds increased with distance, sharply for short distances and then saturating for longer distances. The overall pattern resembled a saturating exponential, similar to model predictions. The overall maximum speed was 1.52 ± 0.14 ms^{–1} (mean ± s.d. across subjects), almost always for the longest distance. We normalized each individual’s peak speed by their own maximum, and found the resulting peak speed vs. distance curves to be scalable into a single normalized curve across subjects. With normalization, the variability (s.d. across individuals) of peak speeds was reduced significantly ($P=1.6\text{\xd7}{10}^{6}$), by $0.07\pm 0.02\phantom{\rule{thinmathspace}{0ex}}\mathrm{m}\phantom{\rule{thinmathspace}{0ex}}{\mathrm{s}}^{1}$ (mean ± s.d.) across all bout distances or about 54% compared to unnormalized. Thus, even though each individual walked at their own pace, that tendency was consistent across all distances. Much of the intersubject variability was reduced by normalizing the peak speeds, revealing a common relationship between peak speed and bout distance.
There was a similarly consistent pattern for walking durations across distances (Figure 5B). Walking durations increased with distance in a slightly curvilinear fashion. Again, we normalized each individual’s durations by the duration for the longest bout ($9.86\pm 0.75\phantom{\rule{thinmathspace}{0ex}}\mathrm{s}$), and found the duration vs. distances to be scalable into a single normalized curve across subjects. With normalization, the variability of durations was also reduced significantly ($P=0.03$), by $0.10\pm 0.13\phantom{\rule{thinmathspace}{0ex}}\mathrm{s}$ across all bout distances, or about 18% compared to unnormalized. Similar to peak speeds, much of the intersubject variability was reduced by normalizing. There was a common and consistent relationship between different walking bouts, similar to model predictions.
The change in peakiness or flatness of speed profiles was indicated by the time spent accelerating, decelerating, or at approximately constant speed (Figure 5B). This was described by rise time, defined as the time to accelerate from 0% to 90% of peak speed, cruise time as the time spent at 90% of peak speed or more, and fall time as the time to decelerate between 90% and 0% of peak speed (Figure 5B). These measures of time increased with bout distance. As a fraction of each bout’s duration, the rise and fall times appeared to take up a greater proportion for shorter bouts, and only a very small proportion was spent at steady speed. Conversely, cruise time took up a greater proportion of the time for longer bouts. These behaviors were consistent with predictions from the EnergyTime hypothesis.
The peak speed was described reasonably well by a saturating exponential (Figure 5A). An ad hoc, leastsquares nonlinear fit to the normalized data yielded a saturating exponential curve
where $v(D)$ is the peak speed as a function of total walking distance $D$, and fitted values were $c}_{v}=1.516\phantom{\rule{thinmathspace}{0ex}}\mathrm{m}\phantom{\rule{thinmathspace}{0ex}}{\mathrm{s}}^{1$ (1.496, 1.536 CI, 95% confidence interval) and $d}_{v}=1.877\phantom{\rule{thinmathspace}{0ex}}\mathrm{m$ (1.798, 1.955 CI), with a goodnessoffit of ${R}^{2}=0.86$. The curve fit shows that there was considerable consistency in maximum speed; even short walking bouts of slow peak speed were still consistent with longer bouts of higher speed.
Similarly, walking duration increased with walking distance (Figure 5B), with a slightly curvilinear relationship. The total walk duration $T(D)$ may be treated as a saturating exponential approaching a straight asymptote, defined ad hoc as distance $D$ divided by preferred steady walking speed plus an offset T_{0}. The curve was of the form
where fitted coefficients were $v}_{T}=1.494\phantom{\rule{thinmathspace}{0ex}}\mathrm{m}\phantom{\rule{thinmathspace}{0ex}}{\mathrm{s}}^{1$ (1.466, 1.521 CI), $T}_{0}=1.470\phantom{\rule{thinmathspace}{0ex}}\mathrm{s$ (1.375, 1.565 CI), and $d}_{T}=0.790\phantom{\rule{thinmathspace}{0ex}}\mathrm{m$ (0.610, 0.970 CI).
We also performed similar analyses on a grass walking surface to test for sensitivity to slightly uneven terrain. An identical set of conditions was collected on short grass outdoors. The fit of peak speed vs. bout distance yielded ${c}_{v}=$ 1.446 ms^{–1}, ${d}_{v}=$ 1.822 ms^{–1} (${R}^{2}=0.85$), and for duration versus distance ${v}_{T}=$ 1.426 ms^{–1}, ${T}_{0}=$ 1.336 s, and ${d}_{T}=$ 0.503 m (${R}^{2}=0.98$). These relationships were quite similar to those obtained on sidewalk.
We next estimated the relationship between human valuation of time and steady walking speed (Figure 6). Here we used an empirical human metabolic power curve (Figure 6A) from literature (Elftman, 1966) to predict how steady walking speed should increase with metabolic value of time ${C}_{T}$ (Figure 6B), and how the energetic cost of transport vs. steady walking speed (Figure 6C) may be regarded in terms of competing costs for Energy and Time. This human power curve was fitted to the model, to facilitate scaling the model’s mechanical energy into human metabolic energy. The optimal steady walking speed emerges from that curve (Srinivasan, 2009), as a function of ${C}_{T}$ (Figure 6A). Based on these crude assumptions, a time valuation of zero yields the same optimal speed ${v}^{*}$ of 1.25 ms^{–1} as min COT, and close to the minimum steady speed (among subjects) of 1.21 ms^{–1} observed here. It is instructive to increment ${C}_{T}$ by multiples of the metabolic equivalent (MET), a standard physiological resting rate of about 1.23 Wkg^{–1} [Watts per kilogram] (Jetté et al., 1990). An increment of 1 or 2 MET yields optimal speeds of 1.54 ms^{–1} or 1.75 ms^{–1}, respectively, quite close to the observed mean and maximum steady speeds (among subjects), respectively (of 1.52 ms^{–1} and 1.75 ms^{–1}). Thus, if the same metabolic power curve were applicable to all, the slowest subject would have valued time at about zero MET, the mean subject at +1 MET, and the fastest at +2 MET. This also suggests that most subjects preferred faster steady speeds than minCOT. Incrementing ${C}_{T}$ yields diminishing returns in speed (Figure 6B), because it is increasingly costly to walk faster. One interpretation afforded by the EnergyTime hypothesis is that there is an effective cost of transport that may be separated into two terms (Figure 6C), Walking and Time: one for the net metabolic cost for walking alone (due to pushoff work), the other a cost of time that lumps the resting rate with (or within) an individual’s ${C}_{T}$. This reveals a tradeoff, where the cost of walking increases with speed, and the cost of time decreases (hyperbolically with speed), such that the two opposing curves (or rather their opposing slopes) determine an optimum. A greater valuation of time adds to this effective cost of transport, equal to the actual metabolic energy plus a subjectively scalable cost of time, per distance traveled. This shows how the effective EnergyTime cost per distance is minimized at higher speeds for greater ${C}_{T}$.
Discussion
We had sought to test whether humans optimize not only metabolic energy but also a valuation of time spent walking. Although the prevailing theory of minimizing the energetic COT explains steady walking, it does not explain shorter walks that lack a steady speed, nor does it readily accommodate individual tendencies toward faster or slower speeds. We found that humans walk bouts of finite distance with a trajectory of speeds varying with distance. These bouts fall within a consistent family of trajectories across subjects, despite individual differences in overall speed or duration. These results are in agreement with a simple mechanistic model of walking, governed by optimization. The findings suggest that humans optimize a combined objective that trades off the energy to arrive at destination against the time it takes to get there.
Each human walking bout consisted of a dynamically varying trajectory of speed with an inverted U shape. Many of these bouts included a period of steady walking, at speed similar to (but usually faster than) the supposed minCOT (Ralston, 1958), but mainly for the longer distances (Figure 4). Shorter bouts of say 10 m or less exhibited a relatively brief peak slower than the typical minCOT speed (Figure 5B). All such bouts also spent substantial time and energy in acceleration and deceleration (Figure 5B). Moreover, short distances such as this are quite ecological, accounting for about half of the daily living walking bouts reported by Orendurff et al., 2008, with acceleration and deceleration potentially accounting for 4–8% of daily walking energy budget (Seethapathi and Srinivasan, 2015). This contrasts with the minCOT hypothesis, which predicts only a single steady speed, and cannot explain how to start or end a bout. People usually walk to a destination of known and fixed distance, for which it is more sensible to minimize the total energy for that distance, rather than energy per distance (minCOT). This is not to dismiss the energy spent for steady walking, which has been wellcharacterized in laboratory settings where the vast majority of published studies have been performed. But in daily living, humans walk a variety of speeds and distances, many too short to be steady. To our knowledge, this is the first study to consider the time and energy spent in brief walking bouts.
Even though there were considerable differences between individuals, each subject was quite consistent within their own walking bouts. Those with a slower or faster peak speed during longer bouts were also consistently so during shorter, nonsteady bouts (Figure 5), as evidenced by the 54% reduced variability after normalizing peak speeds by the longest bout. Moreover, the bouts across all subjects were scalable to a single, selfsimilar family of trajectories (Figure 4). These trajectories were not consistent with a fixed acceleration or deceleration profile (Figure 2A), and instead exhibited a greater peak speed and longer time to that peak with greater distance (Figure 4B). This pattern suggests that there are systematic criteria or principles that govern walking bouts of finite distance. Even though some individuals are faster than others (Figure 4A), they all seem to follow similar principles.
These observations agree with the primary hypothesis that humans optimize for energy and time. The tradeoff between the two, described by the valuation of time ${C}_{T}$ (for human metabolic cost, or ${c}_{T}$ in terms of model work), is readily explained for steady walking. A valuation of zero corresponds to minimizing the gross metabolic cost of transport (Figure 6B), yielding the minCOT speed. (Or equivalently, resting metabolic rate could be treated as part of a running time cost, trading off against the net metabolic rate for walking, as in Figure 6C.) Of course, humans do not always walk the same speed, and faster speeds save time but cost more energy to cover a given distance. One’s actual steady speed must therefore be observed empirically, from which their valuation of time may also be estimated (ranging roughly 0–2 MET, Figure 6B). But what sets time valuation apart is its predictive value for nonsteady walking. The speed trajectory for a fixed bout distance need not even contain a steady portion, and is often dominated by acceleration and deceleration. The EnergyTime model can nevertheless predict an entire family of trajectories across a variety of distances (Figure 3) despite individualspecific step lengths and time valuations. The model also suggests how peak speeds, and time to peak, and durations should increase with distance (Figure 3), similar to human data. All this is based on one free parameter, the individualspecific valuation of time. That valuation may depend on complex physiological and sociopsychological traits, but it nonetheless appears to have predictive value for a given context. Not tested here is the presumption that different contexts, for example changing the saliency of a task or adding time pressure, will also lead to systematic changes in walking bouts. If an individual’s valuation of time can be estimated empirically, our hypothesis provides an operational means of integrating it into a quantitative model.
These predictions are produced by a mechanistic model governed almost entirely by dynamics. The timing comes from the dynamics of pendulumlike walking, and the energetics from the steptostep transition between pendulumlike steps. The steptostep transition requires mechanical work to accelerate and to restore collision losses, such that for short walks it is uneconomical to accelerate quickly to minCOT speed (Figure 2, steady minCOT). An alternative is to accelerate more gradually, but that is costly because of the high peak speed attained (Figure 2, steady accel). The model thus favors an intermediate and smooth acceleration to a slower and continuously varying speed with an inverted U shape. Separate studies have found steptostep transition work to predict human metabolic energy expenditure as a function of step length (Donelan et al., 2002) and changing speed (Seethapathi and Srinivasan, 2015). Here, we have constrained the pendulumlike dynamics so that there is only one physical parameter, step length, which in any case has very little effect on the characteristic shape of speed trajectories (Figure 3). As such, this was set to nominal value and not fitted to data, making the model as predictive as possible. Indeed, the very same model also predicts human compensation strategies for walking on surfaces of uneven height (Darici and Kuo, 2022a; Darici and Kuo, 2022b). Of course, the human body has many degrees of freedom capable of far different motions, but model analysis suggests that pendulumlike walking is the most economical means to move the COM at slow to moderate speeds (Srinivasan and Ruina, 2006), and that pushoff during the steptostep transition is the most economical means of powering such pendulumlike walking (Kuo, 2001). These models are predicated on mechanical work as the major cost, and the COM as the major inertia in the system.
It is instructive to consider what other models might explain or predict our experimental results. We did not explore more complex models here, but would expect similar predictions to result from any model based on pendulumlike walking and steptostep transitions. This includes those cited in the previous paragraph, as well as a family of such models including threedimensional motion (Donelan et al., 2001), knees and biarticular actuators (Dean and Kuo, 2009), and plantarflexing ankles (Zelik et al., 2014). It is also possible that more complex, musculoskeletal models also perform substantial work and expend energy for steptostep transitions, and might therefore agree with the present model. But here, relatively simple principles account for a fairly wide array of predictions (Figure 3), which are unlikely to result from happenstance. We are also unaware of any current hypothesis that could plausibly substitute for the present one. We therefore doubt if alternative models not based on pendulumlike principles could predict or reproduce these results, except with numerous fitted parameters.
This model is optimized with an additional control parameter, for the valuation of time. Time has long been recognized as a factor in the pace of life (Levine and Bartlett, 1984), and in reward and vigor in motor control (Shadmehr et al., 2010). It is typically expressed as a temporal discounting of reward, which appears key to human decision making and the theory of reinforcement learning. Here we expressed it as a tradeoff equivalency between energy and time. This was mainly due to the need for compatibility with our energetics model, but also because neither model nor experiment included an explicit reward to be discounted. We used a simple, linear valuation of time in terms of energy, rather than a nonlinear, exponential or hyperbolic temporal discounting factor (Green and Myerson, 1996). Energy is a physiological cost endemic to life, that is not obviously more or less valuable at different points in time. It is sufficient to predict and explain the present results, and there is currently insufficient evidence to favor a nonlinear cost over our linear valuation. But regardless of the particular formulation, a default valuation of time may be an individualistic trait, generalizable to other tasks such as hand and eye movements (Labaune et al., 2020). Indeed, we have found a similar valuation of time to explain how reaching durations and speed trajectories vary with reaching distance (Wong et al., 2021). Another implication of our model is that humans may incorporate prediction of time within central nervous system internal models. Such models have long been proposed to explain humans predict and adapt their movement trajectories, for example to novel dynamics (Todorov, 2004). If movement duration is also part of human planning, it suggests the ability to predict not only movement trajectories and energetics, but also time. Here such prediction is made operational within a quantitative model.
Valuation of time offers another perspective on minimizing the gross energetic cost of transport. Actual walking tasks are not purely steady, and are probably planned with consideration of what happens at the destination. Long and Srinivasan, 2013 proposed a task to minimize the total energy expended to walk to destination within a more than ample allocation of time. They showed that total energy should be optimized by mixing resting and walking (and running if necessary). Suppose the task is extended to an indefinite duration, where a considerable amount of time is spent resting. The optimal total energy and walking duration may be found by applying our EnergyTime objective with time valuation (${C}_{T}$) equal to zero (Figure 6). Walking faster than optimal would yield more time to rest, but at a greater total energy cost for walking. Walking slower would cost less energy for the walking motion alone, but at a greater total cost due to less time available to rest. After all, ${C}_{T}$ is the energy one is willing to expend to save a unit of time, and the resting rate is the energy expended to rest for a unit of time. This may seem like a trivial restatement of the minCOT hypothesis, but it differs in two important ways. First, it can predict both the duration of walking and the entire speed trajectory, even for short bouts where there is no steady portion. Second, it considers how valuable time is at the destination. Minimizing the gross cost of transport is most sensible for maximizing the survivable range distance (Srinivasan, 2009), which may not be a concern in modern life where survival rates are high, walks short, and calories plentiful. Rather, it may be a sensible default to value time at close to the resting rate (particularly for long walks), and then to vary the valuation dependingon context. One might thus rush toward a longlost friend or hurry in a big city, if the time spent at destination is far more valuable than resting (Bornstein, 1979; Levine and Bartlett, 1984; Li and Cao, 2019). Similarly, we do not consider walking slowly to be a waste of energy per distance, but rather a waste of time. Even then, there are cases when humans might wish to waste time, for example to avoid an odious task, according to the expression ‘the slow march to the gallows’.
The consistency of individual walking trajectories may have practical implications. Although walking speed is used as a clinical indicator of mobility, it is difficult to standardize (Middleton et al., 2015), because evaluations may be confined to the length of the available walkway, which may be too short (e.g. less than 10 m) for a steady speed to be reached. But given the time to walk a fixed distance, it may be possible to predict the duration and steady speed for another distance, referenced from a universal family of walking trajectories. We have identified one such family that applies to healthy young adults with pendulumlike gait. We do not know whether that family also applies to older adults, who prefer slower steady speeds and expend more energy to walk the same speed (Malatesta et al., 2003). Perhaps an agerelated valuation of time might explain some of the differences in speed. Of course, some clinical conditions might be manifested by a deviance from that family, perhaps in the acceleration or deceleration phases, or in how the trajectories vary with distance. If quantified, such deviance might prove clinically useful. The methodology employed here does not require specialized equipment beyond inertial measurement units, and the characterization of distancedependent speed trajectories can potentially provide more information than available from steady speed alone.
The EnergyTime hypothesis could be tested by further inquiries. We have thus far regarded the valuation of time as a relatively fixed parameter for each subject. That valuation is likely influenced, and therefore testable, by many contextual factors, including physiological and sociopsychological variables and task constraints. For example, caffeine intake, feeding status (e.g. Taylor and Faisal, 2018), or monetary reward could be used make time more valuable as a tradeoff against energy. Conversely, energy may be helpful for assessing the valuation of time (or temporally discounted reward), which is not easy to measure other than indirectly. (Energy is also a universal currency, because all animals use energy, whereas only some use money.) Walking has a wellcharacterized physiological energy cost, and could serve as a useful tradeoff against time or reward. The hypothesized optimal gait is the point at which the costs of energy and time have equal and opposite slopes (i.e., partial derivatives) with respect to an independent variable such as speed (e.g. Figure 6C), carried load, or incline. There are thus a variety of opportunities to manipulate the energetic cost of walking, as a means to assess the proposed valuation of time.
There are a number of limitations to this study. Although we tested model predictions in terms of speed trajectories, we did not measure mechanical work or metabolic energy expenditure in human subjects, which would provide greater insight regarding the proposed tradeoffs against time. We also did not evaluate each individual’s metabolic cost of transport vs. speed, which would reveal more precise differences between the minCOT speed and the actual selfselected speed. Nor did we evaluate gait kinematics or kinetics, which may be helpful for detailing other ways that walking bouts vary with distance. The simple walking model also only includes a crude representation of steptostep transitions, which we have crudely estimated to account for as much as 70% of net metabolic cost in steady walking (Donelan et al., 2002). We did not include other factors such as forced leg motion and step length modulation (Doke et al., 2005, up to say 33% of cost) that likely also affect energetic cost, and could therefore be used to test the valuation of time. Nor did we include factors such stability (Bauby and Kuo, 2000; Donelan et al., 2004; Rebula et al., 2017) and threedimensional motion (Donelan et al., 2001), despite being part of our previous models, because we believe them to contribute little to the present task. In fact, because the optimal control model successfully completes the walking task, its feedforward motion attains implicit stability, which could reduce (but not eliminate) the need for feedback stabilization (Darici and Kuo, 2022a; Darici and Kuo, 2022b). We also did not include an explicit reward, which could facilitate assessment of energy and time in terms of other tradeoffs such as money or food. In fact, the EnergyTime hypothesis should be regarded as a subset of the many factors that govern human actions, rendered here in a simple but quantitative form.
Conclusion
Humans appear to select walking speed dynamically to minimize a combination of energy and time expenditure. This is both compatible with and extends the traditional hypothesis that humans minimize gross energy expenditure per unit distance. We found it more accurate to minimize the total cost of a walking bout, due to the ability to predict an entire speed trajectory, with the optimal steady speed as an emergent property. By including a cost for time expenditure, we introduce a quantitative and operational means to make walking models compatible with the study of movement vigor. Tasks may also be broadened beyond walking, to include consideration of the reward to be gained or further energy to be expended once the destination is reached. Walking may thus be integrated into broader questions of how and why humans take the actions they do. As a modification to the traditional adage about money, we suggest that ‘Time is energy’.
Methods
We experimentally tested how human walking speed varies with walking distance. The speed trajectories observed from human subjects were compared against predictions from the EnergyTime hypothesis and against the minimumCOT speed. To formulate the hypothesis and make quantitative predictions, we expressed it as an optimal control problem including both energy and time. We first state the hypothesis for human walking, and describe how it is adapted for a simple walking model to yield predicted speed trajectories. This is then followed by description of the experiment regarding human walking speed, and finally an analysis of steady speed as a property of the model.
Walking model
We use the ‘simplest walking model’ (Kuo, 2002) to operationalize this optimization problem (Figure 7A). The model treats the stance leg as an inverted pendulum and requires mechanical work to power the gait. The body center of mass (COM) is modeled as a point mass supported by the stance leg, so that each pendulumlike step follows an arc, which itself requires no energy input. Work is performed during the steptostep transition (Figure 7B), to redirect the COM velocity from forwardanddownward the end of one arc, and forwardandupward at the beginning of the next. This is accomplished most economically with an active, impulsive pushoff along the axis of the trailing leg, immediately followed by an impulsive, dissipative collision between the rigid leading leg and ground. In steady gait, the optimal pushoff restores the collision losses, with mutually canceling impulses of equal magnitude. Speeding up is a matter of a greater pushoff than collision, and a net increase in COM velocity during the steptostep transition (Figure 7C). Positive and negative work are proportional to the square of the pushoff and collision impulses, respectively (Kuo, 2002), so that speeding up also dissipates less collision energy than steady gait. Slowing down is the same in reverse, with collisions exceeding pushoffs. This model predicts how steptostep transition work for steady walking should increase as a function of step length and step width (Donelan et al., 2002). The model mainly predicts mechanical work for pushoff, which appears to be a proportional predictor of the majority of human metabolic energy during steady walking (Donelan et al., 2002). That work also yields a mechanical cost of transport that varies curvilinearly with steady speed, similar to the empirical metabolic curve (Figure 1a; Ralston, 1958). There are of course other contributions to the metabolic cost of walking such as to move the swing leg (Kuo, 2001), but of smaller magnitude than steptostep transitions, which are to be tested alone for their predictive value. Details of this model have been described in greater detail previously (Darici et al., 2020; Kuo, 2002), and are recounted only briefly here.
A walking bout consists of a sequence of $N$ steps, starting and ending at rest. It may be described by the discrete sequence of body speeds v_{i} ($i=1,2,\mathrm{\dots},N$), each equal to the distance traveled for step $i$ divided by that step’s time duration ${\tau}_{i}$. The model begins at rest in an upright position (Figure 7D), and is set into motion by a forward initiation impulse acting on the pelvis. In humans, the torso can serve as an inertia that the hip muscles can act against, but for simplicity this action is represented as a translational impulse at the pelvis, summarized by the associated positive work u_{0}. The total positive work performed by the model consists of the work from initiation and the successive pushoffs, a sequence u_{i} ($i=0,1,\mathrm{\dots},N$). There is also a corresponding sequence of dissipative collision impulses by the leading leg, and a dissipative gait termination to end at upright.
The steptostep transition starts just before leading leg ground contact contact, and consists of a perfectly impulsive pushoff from the trailing leg, followed in immediate succession by a perfectly inelastic and impulsive collision of the leading leg with ground. The COM velocity at the end of one stance phase is ${v}_{i}^{}$, directed forward and downward according to the pendulum arc. Mechanical work is performed only during the steptostep transition, with a succession of ideal impulses. First is positive pushoff work from the trailing leg, directed from its foot to the COM, and second is a perfectly inelastic heelstrike collision of the leading leg with ground, directed from the leading foot to the COM. For brevity, the equations presented here use dimensionless versions of quantities, with body mass $M$, gravitational acceleration $g$, and leg length $L$ as base units. The pushoff work is denoted u_{i} (in units of massnormalized work), and the pushoff and collision sequence act to redirect the COM velocity to ${v}_{i}^{+}$ at the beginning of the next stance phase, directed forward and upward according to the next pendulum arc. Using impulsemomentum, the steptostep transition is described by
where $2\alpha $ is the interleg angle (Figure 7A). There is no work performed during the passive, inverted pendulum phases, and so the steptostep transition is responsible for all energy inputs (u_{i}) and energy losses (from collisions).
The dynamics of an inverted pendulum describe all of the other motion in the system, consisting of the falling of one inverted pendulum toward the steptostep transition, and the rising of the next inverted pendulum toward midstance. These dynamics determine the respective velocities and timing of these respective instances. The velocities may be found through conservation of energy:
The step time ${\tau}_{i}$ is defined as the time for the stance leg angle $\theta $ to move between successive midstance instants, and the corresponding velocities from v_{i} to ${v}_{i+1}$. It may be regarded as the sum of a time ${\tau}_{i}^{}$ from midstance to the steptostep transition, and then the time ${\tau}_{i}^{+}$ from the steptostep transition until next midstance. Using the linearized dynamics, the dimensionless time ${\tau}_{i}^{}$ of step $i$ is
The other time ${\tau}_{i}^{+}$ is
For comparison with experiment, we also defined an average (as opposed to midstance) speed for each step $i$ as the step length divided by the step time between midstance instances,
The trajectory of this body speed is plotted for different walking bouts, for both model and experiment. The equations for body speed and step time are summarized as constraints $f$ and $g$ below.
We chose nominal parameters to correspond to typical human walking. A person with body mass $M$ 70 kg and leg length $L$ of 1 m may typically walk at 1.25 ms^{–1}, with step length of 0.68 m and step time of 0.58 s, and corresponding fixed constant value $\alpha =0.35$. The step length was also varied in parameter sensitivity studies. Using dynamic similarity, parameters and results are reported here either in SI units, or in normalized units with body mass $M$, gravitational acceleration $g$, and $L$ as base units.
Optimal control formulation
We applied optimal control to the model for short walk bouts of varying distance (Figure 7). In humans, both positive and negative work appear to cost positive metabolic energy with different proportionalities (Margaria, 1976). In the model, we assess a cost only for positive work, because the net work of a level walking bout is zero. Minimizing positive work thus also implicitly minimizes the negative work, as well as metabolic cost of any proportionality. The pushoffs have a onetoone relationship with the speeds, and so either pushoffs or speeds can can describe the trajectory. For the model, the goal is to minimize an objective function ${J}_{\mathrm{model}}$ comprising the total positive work for the walking bout, plus the cost for the time duration:
where the coefficient ${c}_{T}$ is the model’s valuation of time in terms of work, and equal to the mechanical work the model is willing to spend to save a unit of time. It is treated as proportional to the human’s valuation ${C}_{T}$ for metabolic energy per time.
This objective is applied as follows. The total distance $D$ of a walking bout may be achieved by taking an appropriate number of steps $N$. The walking trajectory is described by a discrete sequence of speeds v_{i} (step $i=1,2,\mathrm{\dots},N$), starting and ending from standing at rest, given a standard step length. The corresponding control actions include the initiation impulse and the pushoff impulses, for a total of $N+1$ actions u_{i} ($i=0,1,2,\mathrm{\dots},N$). Using these variables, the model’s objective is thus
for the optimization problem
where the model begins and ends at rest, and walking dynamics constrain how the speed and duration of the next step depend on the current step’s speed and pushoff (functions $f$ and $g$ detailed above). Note there are actually $N+1$ controls, consisting of the initiation input u_{0} and the $N$ steptostep transition pushoffs $({u}_{1},{u}_{2},\mathrm{\dots}{u}_{N})$.
The time valuation ${c}_{T}$ is treated as an unknown but constant coefficient. Greater ${c}_{T}$ is expected to yield faster walking bouts, with experimental data used to determine an appropriate range of values. Within a fixed experimental context, we expect ${c}_{T}$ to be constant. We found values of ${c}_{T}$ ranging 0.006–0.06 $M{g}^{1.5}{L}^{0.5}$ to yield speeds approximately similar to subjects.
Step lengths were examined with a nominal fixed step length of 0.68 m, and sensitivity analyses performed with fixed lengths of 0.59 m and 0.78 m, and varying lengths following the human preferred step length relationship $s={v}^{0.42}$(Grieve, 1968). We have previously proposed that steptostep transitions account for most of the metabolic cost of human walking (say, up to 70% Donelan et al., 2002), and that forced swing leg motions to modulate step length also contribute a nonnegligible (Kuo, 2001), but lesser cost (say, up to 33% Doke et al., 2005) that is neglected here. This is not to dismiss this and other costs of locomotion, but merely to hypothesize that steptostep transitions will still dominate in transient walking bouts. The failure to include other costs, if sufficiently critical, should cause the model to make poor predictions.
We also considered two alternative hypotheses. One was that walking occurs almost entirely at the optimal steady speed. Termed the steady minCOT hypothesis, the goal is to walk at the minCOT speed ${v}^{*}$, or close to it, as much as possible. This is accomplished by minimizing deviations from ${v}^{*}$ throughout the bout, with objective
subject to the same constraints as the EnergyTime hypothesis. This objective is expected to cause the model to accelerate immediately from rest to ${v}^{*}$, then remain at that steady speed, and then finally decelerate immediately back to rest. The second alternative was a steady acceleration hypothesis, which contrasts with minCOT’s immediate high acceleration. Here, the acceleration is made as gradual as possible, albeit at the expense of a higher peak speed needed to travel the same distance and duration. Both of these alternatives help illustrate how different speed trajectories requires differing amounts of mechanical work, to be compared against the work produced by the EnergyTime hypothesis.
Model predictions were produced using computational optimization. Optimal control was computed using the JuMP optimization package for the Julia language (Dunning et al., 2017), formulated as a discrete collocation problem, minimized by nonlinear programming (Ipopt). Walking bouts were conducted for $N$ ranging 1–20 steps. The resulting trajectories were condensed into a scalable, selfsimilar family of speed trajectories.
Experimental methods
We tested the model predictions by experimentally measuring the speed profiles of healthy adults walking a series of short distances, ranging about 2–20 steps. Subjects ($N=10$, 6 male and 4 female, 24–38 yr) were instructed to walk at a comfortable speed in ten distance conditions, starting from standing at one numbered marker on the ground, and ending at another as requested by the experimenter. After each trial, there was a brief waiting interval of about 15–30 s, to reduce interference between successive trials and to avoid any incentive to rush through trials. The walking surface was a level sidewalk. The numbered markers were separated by distances of 1.1, 1.7, 2.2, 2.8, 3.3, 3.8, 5.1, 7, 9.1, and 12.7 m. Subjects were provided with a simple task upon reaching the target: They were provided a pointer stick and instructed to walk to and touch the pointer to the target marker. This was intended to provide a context for the task, reflecting the fact that humans often walk to a particular destination to accomplish a task. Each distance condition was conducted a total of four times in two pairs of outandback trials, with the distances in random order. There were therefore a total of 400 trials, from 10 subjects walking 10 distances, each four times.
Walking speeds were measured from footmounted inertial measurement units (IMUs). These were used to compute the spatiotemporal trajectory of each foot in 3D, which was then processed to yield forward walking speed for the body per step. Each IMU (Opal sensors, APDM Inc, Portland, Oregon) was placed on the top of each foot, taped to the outside of the shoe. The recorded data of linear acceleration and angular velocity data were integrated using a previously described algorithm (Rebula et al., 2013) to yield foot trajectories. Briefly, the algorithm detects footfalls as instances in time when the foot is momentarily at rest on the ground, as defined by thresholds for acceleration and angular velocity. The footfall instance was defined as the midpoint of the belowthreshold interval, and used to correct the integrated foot velocity (from gravitycorrected inertial accelerations) to zero, thus reducing IMU integration drift. The footfalls were also used to segment data into discrete strides, from which speed and length of each stride was calculated. (Subjects also wore another IMU on a waist belt, the data from which was used to demarcate the trials, but not for any further quantitative analysis.)
There were a few other analysis adjustments required to produce forward walking data. The absolute position and compass heading of the IMUs were unknown, yielding independent foot trajectories with no relation to each other. However, the experimental conditions called for forward walking for a known distance, so we rotated each foot path to align them into a single forward direction. We also assumed that both feet travelled approximately the same distance for each walk, and translated and rescaled the start and end points to match each other, to yield a processed positiontime graph of the two feet (see representative data in Figure 8). We also devised a definition for the starting and ending times for each trial based on IMU data. Humans initiate their gait by shifting their weight before moving the feet (Mann et al., 1979), so that the footfall threshold defined above may not detect the actual gait initiation. We therefore defined a rough approximation to gait initiation and termination, starting before and ending after threshold crossing, by an amount equal to half the average belowthreshold time during walking. This adjustment may be incorrect compared to actual weight shift by several tens of milliseconds. The experiment is mainly concerned with speed profiles over time on the order of several seconds. The accuracy of the experiment can thus tolerate small errors in detecting gait initiation or termination.
Finally, the body’s walking speed and length of each step were calculated as follows (Figure 8). The trajectory of each foot’s strides were found to cross each other, approximating the time in midstance when one foot passes by the other. These points of intersection were used to define step length as the spatial distance between intersections, and step time as the temporal difference between intersections. These were used to determine the average speed for each step (‘body speed’), defined as the step length divided by step time ending at each intersection. This assumes that the body moves as much as the feet between midstance instances. These discrete, stepaveraged data were used to produce trajectories of body speed for each bout (Figure 8, inset), without regard to continuoustime undulations in velocity for the body center of mass. For comparison with these data, similar discrete body speeds and times were computed from model predictions.
We used these data to test the model predictions. We examined how human speed profiles varied with bout distance, and exhibited more rounded peaks for shorter bouts and flatter ones for longer ones. We tested for selfsimilarity by scaling the profiles by speed and time and performing statistical tests regarding peak speeds and walking durations. We tested whether a saturating exponential describes the increase in peak speed with bout distance (${R}^{2}$; test 95% confidence interval of parameters not including zero). Expecting a selfsimilar shape for the peak speed vs. distance relationship, we scaled the curves by peak speed and tested for a single exponential. We tested selfsimilarity in terms of a reduction of variability in peak speed (standard deviation across subjects) for each condition, comparing nonnormalized to normalized peak speeds (rescaled to mean overall peak speed) with paired ttest. We examined the walking durations as a function of bout distance, and also tested selfsimilarity by significant reduction in standard deviations across subjects, comparing nonnormalized to normalized data (rescaled to mean longest duration) with paired ttest. We also described walking durations in terms of rise and fall times (between 10% and 90% of peak speed).
Prior to the experiment, subjects provided informed consent as approved by the University of Calgary Conjoint Health Research Ethics Board (REB211497). Preestablished exclusion criteria included significant health or other conditions that preclude ability to walk on uneven terrain or moderate hiking trails; no prospective participants were excluded. Subjects were recruited from the community surrounding the University of Calgary; the city has a moderately affluent population of about 1.4 M, with a developed Western culture. The experiment was performed once.
Effect of valuation of time on steady walking speed
We performed an additional analysis to consider how the hypothesized energetic value of time may affect human steady walking speeds (Figure 6). This requires a valuation of time in terms of human metabolic energy rather than the model’s mechanical work, and a consideration of longer walking bouts where steady walking dominates. To empirically quantify human cost as a function of speed, we fitted the model’s steady mechanical work rate to human net metabolic power reported by Elftman, 1966, with a resting power adjusted to agree with the optimal steady speed of 1.25 ms^{–1} reported by Ralston, 1958. The model was of the form
where $a$, $b$ and $d$ are empirical coefficients, and $n$ is a model constraint. The exponent $n$ is not critical, and values ranging 2–4 are sufficient to describe the increase. However, we used a value of $n=3.42$ as predicted by the simple model for humanlike walking (Kuo, 2002). For metabolic power in, the empirical coefficients are $a=4.90\phantom{\rule{thinmathspace}{0ex}}\mathrm{W}\mathrm{k}{\mathrm{g}}^{1}$, $b=1.16\phantom{\rule{thinmathspace}{0ex}}\mathrm{m}{\mathrm{s}}^{1}$, and $d=1.56\phantom{\rule{thinmathspace}{0ex}}\mathrm{W}\mathrm{k}{\mathrm{g}}^{1}$ (${R}^{2}=0.99$). The yintercept may be regarded as a resting rate, at 1.73 Wkg^{–1} (Figure 6A). The resulting cost is therefore proportional to the model’s mechanical work, while matching well with human metabolic power and optimal steady speed data. The curve may be expressed as cost of transport by dividing power by speed, $\dot{E}/v$.
We then used our own walking data to estimate the human valuation of time. We used the peak walking speeds from the longest walking bout as indicator of steady speed. These were compared to the steady speed predicted by the metabolic cost curve with an added variable, the metabolic valuation of time ${C}_{T}$. The EnergyTime curve was converted to cost of transport, and then minimized to yield optimal speed. This is equivalent to taking the limit of the EnergyTime objective as function of increasing distance, thus making the costs of starting and ending a walking bout small. The result predicts that steady speed will increase approximately with the cube root of ${C}_{T}$ (Figure 6B). This curve was thus used to estimate ${C}_{T}$ for experimentally observed range of steady speeds. It was also used to estimate the effective cost of transport, including the valuation of time, as a function of speed (Figure 6C). This cost of transport may further be regarded as the sum of separate costs for Walking and Time (Figure 6C), where Walking refers to the cost of transport due to pushoff work alone, and Time refers to the cost of transport due to the ${C}_{T}$ term alone.
Data availability
Data (Carlisle and Kuo, 2023, https://github.com/kuolab/short_walk_experiment) and code (https://github.com/kuolab/simplelocomotionmodel, copy archived at Kuo, 2023) for this study are in publiclyaccessible archives.
References

Gait speed as an incremental predictor of mortality and major morbidity in elderly patients undergoing cardiac surgeryJournal of the American College of Cardiology 56:1668–1676.https://doi.org/10.1016/j.jacc.2010.06.039

Arbeitsphysiologische studienPflüg Arch Ges Physiol 215:291–328.https://doi.org/10.1007/BF01731318

Active control of lateral balance in human walkingJournal of Biomechanics 33:1433–1440.https://doi.org/10.1016/s00219290(00)001019

The pace of life: RevisitedInternational Journal of Psychology 14:83–90.https://doi.org/10.1080/00207597908246715

Effects of obesity and sex on the energetic cost and preferred speed of walkingJournal of Applied Physiology 100:390–398.https://doi.org/10.1152/japplphysiol.00767.2005

Elastic coupling of limb joints enables faster bipedal walkingJournal of the Royal Society, Interface 6:561–573.https://doi.org/10.1098/rsif.2008.0415

Mechanics and energetics of swinging the human legThe Journal of Experimental Biology 208:439–445.https://doi.org/10.1242/jeb.01408

Mechanical and metabolic determinants of the preferred step width in human walkingProceedings. Biological Sciences 268:1985–1992.https://doi.org/10.1098/rspb.2001.1761

Mechanical work for steptostep transitions is a major determinant of the metabolic cost of human walkingThe Journal of Experimental Biology 205:3717–3727.https://doi.org/10.1242/jeb.205.23.3717

Mechanical and metabolic requirements for active lateral stabilization in human walkingJournal of Biomechanics 37:827–835.https://doi.org/10.1016/j.jbiomech.2003.06.002

JuMP: A modeling language for mathematical optimizationSIAM Review 59:295–320.https://doi.org/10.1137/15M1020575

Biomechanics of muscleThe Journal of Bone & Joint Surgery 48:363–377.https://doi.org/10.2106/0000462319664802000017

Fuel oxidation in relation to walking speed: Influence of gradient and external loadEuropean Journal of Applied Physiology 110:515–521.https://doi.org/10.1007/s0042101015236

Gait patterns and the speed of walkingBioMed Engng 3:119–122.https://doi.org/10.1007/9781349048359_32

A simple model of bipedal walking predicts the preferred speedstep length relationshipJournal of Biomechanical Engineering 123:264–269.https://doi.org/10.1115/1.1372322

Energetics of actively powered locomotion using the simplest walking modelJournal of Biomechanical Engineering 124:113–120.https://doi.org/10.1115/1.1427703

Energetic consequences of walking like an inverted pendulum: Steptostep transitionsExercise and Sport Sciences Reviews 33:88–97.https://doi.org/10.1097/0000367720050400000006

SoftwareSimple locomotion model, version swh:1:rev:fc46ca7f88d6f91ec43bb77f0d8809df2eb73ee8Software Heritage.

Vigor of reaching, walking, and gazing movements: On the consistency of interindividual differencesJournal of Neurophysiology 123:234–242.https://doi.org/10.1152/jn.00344.2019

Minimum cost of transport in Asian elephants: Do we really need a bigger elephant?The Journal of Experimental Biology 215:1509–1514.https://doi.org/10.1242/jeb.063032

Pace of life, punctuality, and coronary heart disease in six countriesJournal of CrossCultural Psychology 15:233–255.https://doi.org/10.1177/0022002184015002009

Moving at the speed of life: How life PACE influences temporal reasoningMetaphor and Symbol 34:158–166.https://doi.org/10.1080/10926488.2019.1649839

Walking, running, and resting under time, distance, and average speed constraints: Optimality of walkrunrest mixturesJournal of the Royal Society, Interface 10:20120980.https://doi.org/10.1098/rsif.2012.0980

Energy cost of walking and gait instability in healthy 65 and 80yroldsJournal of Applied Physiology 95:2248–2256.https://doi.org/10.1152/japplphysiol.01106.2002

The initiation of gaitThe Journal of Bone and Joint Surgery. American Volume 61:232–239.https://doi.org/10.2106/0000462319796102000011

Effects of age and physical activity status on the speedaerobic demand relationship of walkingJournal of Applied Physiology 73:200–206.https://doi.org/10.1152/jappl.1992.73.1.200

Walking speed: The functional vital signJournal of Aging and Physical Activity 23:314–322.https://doi.org/10.1123/japa.20130236

How humans walk: Bout duration, steps per bout, and rest durationJournal of Rehabilitation Research and Development 45:1077–1089.https://doi.org/10.1682/jrrd.2007.11.0197

Energyspeed relation and optimal speed during level walkingInternationale Zeitschrift Fur Angewandte Physiologie, Einschliesslich Arbeitsphysiologie 17:277–283.https://doi.org/10.1007/BF00698754

The stabilizing properties of foot yaw in human walkingJournal of Biomechanics 53:1–8.https://doi.org/10.1016/j.jbiomech.2016.11.059

BookEnergetics of walking, Ch. 5 (3rd edition)Philadelphia: Lippincott Williams & Wilkins.

Temporal discounting of reward and the cost of time in motor controlThe Journal of Neuroscience 30:10507–10516.https://doi.org/10.1523/JNEUROSCI.134310.2010

Contributions of metabolic and temporal costs to human gait selectionJournal of the Royal Society, Interface 15:20180197.https://doi.org/10.1098/rsif.2018.0197

Optimality principles in sensorimotor controlNature Neuroscience 7:907–915.https://doi.org/10.1038/nn1309

Gaitspecific energetics contributes to economical walking and running in emus and ostrichesProceedings. Biological Sciences 278:2040–2046.https://doi.org/10.1098/rspb.2010.2022

Fuel oxidation during human walkingMetabolism 54:793–799.https://doi.org/10.1016/j.metabol.2005.01.024

The role of series ankle elasticity in bipedal walkingJournal of Theoretical Biology 346:75–85.https://doi.org/10.1016/j.jtbi.2013.12.014
Decision letter

Gordan J BermanReviewing Editor; Emory University, United States

Joshua I GoldSenior Editor; University of Pennsylvania, United States

Sam BurdenReviewer; University of Washington, United States
Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.
Decision letter after peer review:
Thank you for submitting your article "Humans dynamically optimize walking speed to save energy and time" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Joshua Gold as the Senior Editor. The following individual involved in the review of your submission has agreed to reveal their identity: Sam Burden (Reviewer #1)
The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.
Essential revisions:
1) The title asserts that "humans dynamically optimize walking speed to save energy and time". Directly substantiating this claim would require independently manipulating the (purported) energy and time cost of walking for human subjects, but these manipulations are not undertaken in the present study. What the Results actually report are two findings: a. (simulation) minimizing a linear combination of energy and time in an optimal control problem involving an invertedpendulum model of walking bouts that (i) start and stop at rest and (ii) walk at constant speed yields a gentlyrounded speedvstime profile (Figure 2A); b. (experiment) human subject walking bouts that started and stopped at rest had selfsimilar speedvstime profiles at several bout lengths after normalizing by the average duration and peak speed of each subject's bouts (Figure 4B). If the paper established a strong connection between (a.) and (b.), e.g. if speedvstime trajectories from the simulation predicted experimental results significantly better than other plausible models (such as the 'steady minCOT' and 'steady accel' models whose trajectories are shown in Figure 2A), this finding could be regarded as providing indirect evidence in support of the claim in the paper's Title. Accordingly, it would be more accurate to assert 'brief human walking bouts look like trajectories of an invertedpendulum model that minimize a linear combination of energy and time' (although this is obviously too wordy). But unfortunately, the connection between (a.) and (b.) is only discussed qualitatively, and the other plausible models introduced in the Results are not revisited in the Discussion. The representative 'steady minCOT' trace in Figure 2A may be a real contender with the 'EnergyTime' trace for explaining the experimental results in Figure 4, but this candidate is rejected at the end of the thirdtolast paragraph in the "Model Predictions" subsection of Results based on the vague rationale that is never revisited. Reexamining and discussing these points will be important in a subsequent revision.
2) An additional limitation of the approach not discussed in the manuscript is that a fixed step length was prescribed in the simulations. The "Optimal control formulation" subsection in "Methods" summarizes the results of a sensitivity analysis conducted by varying the fixed step length, but all results reported here impose a constantsteplength constraint on the optimal control problem. Although this is a reasonable modeling simplification for steadystate walking, it is less wellmotivated for the walking bouts considered here that start and stop at rest. For instance, the representative trial from a human subject in Figure 8 clearly shows initiation and termination steps that differ in length from the intermediate steps (visually discernable via the slope of the dashed line interpolating the black dots). Presumably different trajectories would be produced by the model if the constantsteplength constraint were removed. It is unclear whether this change would significantly alter predictions from either the 'EnergyTime' or 'steady minCOT' model candidates. While the reviewers agreed that this change would entail substantial work that may be out of the scope of the present paper, they did agree that it is important to at least discuss this limitation.
3) Related to the above, the reviewer thought that the last two sentences in the abstract are too strong and should be rephrased to be more accurate and defensible.
4) Please explain the rationale for the competing objective functions (constant acceleration, in particular). The minCoT objective appears supported by literature, but it was clearly less plausible for transient gait than the energytime objective. The constant acceleration and minCoT objectives did not seem like terribly challenging competing hypotheses to refute. Perhaps the authors can test or comment on objectives that produce realistic families of speed profiles to the energytime hypothesis.
5) The rationale for defining rise and fall times based on 90% of the peak speed is unclear. The experimental acceleration profiles appear slightly steeper than the deceleration profiles, up to ~50% of the max speed. Consequently, the interpretation of rise and fall times may depend on this threshold. Can the authors justify their threshold or report the sensitivity of their conclusions to alternative thresholds?
6) As the authors have a strong history of using simple models to study the energetics of gait, a (possibly speculative) discussion on how individual differences in specific features of locomotion (stabilization, leg swing, etc…) may have on the estimated valuations of time may improve readers' evaluation of the immediate generalizability of the paper's conclusions.
7) In accordance with eLife policies, the authors need to release their data and simulation code, as well as provide details about their experimental apparatus.
8) The manuscript would benefit from careful proofreading prior to publication – the reviewers caught a number of typographical and grammatical errors.
Reviewer #1 (Recommendations for the authors):
The manuscript would benefit from careful proofreading prior to publication – I caught a number of typographical and grammatical errors.
Title
As discussed in my public review, I think the claim asserted in the title is too strong and should be rephrased to be more accurate/defensible.
Abstract
I think the last two sentences in the abstract are too strong and should be rephrased to be more accurate/defensible.
Introduction
– I think it is important to clarify that the results summarized in the opening sentences pertain to laboratory studies. Particularly given that a key motivator for the present study is realworld relevance, it would be valuable to clarify what is known from the lab vs in the wild.
– The (Hoyt and Taylor, 1981) cite appears to only report data on horses (not “other animals”) – please confirm/clarify.
Results
Model predictions
– It was unclear to me on my first couple of readthroughs that there is a constant steplength constraint enforced in the optimization problem – this constraint should be clearly specified in the display equation that contains the plainlanguage formulation of the optimal control problem and/or the adjacent text.
– Perhaps refer to the agent here as a “walker” rather than “human”, both for clarity and so as to not anthropomorphize the model.
– The concluding paragraph of this subsection contains a lot of weasel words and should be revised. https://en.wikipedia.org/wiki/Weasel_word
Figure 2
– Am I understanding correctly that the EnergyTime trace has a smaller integral in (B) than the other two traces? If so, I have two suggestions:
1. Include a subfigure/overlay showing the integral of Work for the three traces in (B) to emphasize this point;
2. In the sentence “Energytime hypothesis predicts the least total work, whereas steady minCOT and steady acceleration require more overall work.”, the two clauses are redundant, and I think Time should be capitalized for consistency with the rest of the document.
Figure 4
– Plots with many overlapping opaque lines are difficult to read and slow to render. I think the ‘One subject, all trials’ traces are fine asis, but I suggest instead one of the following revisions to the ‘All subjects, all trials’ traces:
1. Adjust trace opacity (rather than lightness) so that overlaps visually convey density – this makes the figure easier to read, but further degrades rendering speed;
2. Replace traces with filled region(s) / trace(s) illustrating percentiles of data spread, e.g. interquartile [25%,75%] (and/or alldata [0%,100%]) fill with median (50%) (and/or quartile (25%,75%)) trace(s).
Figure 5
– Are the ‘Individuals’ traces shown pre or postnormalization?
– The contrast between ‘raw’ and ‘normalized’ whiskers is difficult to discern.
– The visualization of Fall/Cruise/Rise times in (B) is fantastic!
Figure 6
– I found the phrase “virtual cost” confusing – isn’t the point of this paper to show that energy and time costs are on equal footing? I think the sentence is clearer if “is a virtual cost and” is removed.
Discussion
– It wasn’t clear to me why “time spent at destination is far more valuable than resting” in a big city – it would be helpful if the authors could clarify this claim with reference to the underlying literature.
– The idea in the penultimate paragraph to further test the EnergyTime hypothesis by manipulating the energetic cost of walking is clever! I hope the authors pursue this
– I suggest replacing "should" (which connotes authority) with "presumably" or "may" (to convey that this is a hypothesis, not an obvious truth).
Conclusion
– The phrase "we found it more general" is vague and (in my reading) not wellsubstantiated by the results. I suggest revising/rephrasing.
– It might be fun to cite the (purported) origin of the "time is money" aphorism.
https://en.wikipedia.org/wiki/Time_is_money_(aphorism)
Methods
– I appreciated the care that clearly went into the simulation and experimental study designs and the clear and concise writing in this section. In particular, I appreciated the discussion of tolerable errors in the timing of gait initiation and termination.
– It may be better to refer t“ "stepaveraged body speed" – I was confused for a long time during my first read by the coarse and irregular time quantization in the results figures. Only once I read and internalized the simulation and experiment methods did I understand (i) the cleverness of the study design and (ii) how to interpret the figures.
– I was absolutely delighted to learn about JuMP! Thank you for teaching me – I commend your use of opensource tools and hope you’ll make your source code available to the community.
Reviewer #2 (Recommendations for the authors):
Please explain the rationale for the competing objective functions (constant acceleration, in particular). The minCoT objective appears supported by literature, but it was clearly less plausible for transient gait than the energytime objective. The constant acceleration and minCoT objectives did not seem like terribly challenging competing hypotheses to refute. Perhaps the authors can test or comment on objectives that produce realistic families of speed profiles to the energytime hypothesis.
The rationale for defining rise and fall times based on 90% of the peak speed is unclear. The experimental acceleration profiles appear slightly steeper than the deceleration profiles, up to ~50% of the max speed. Consequently, the interpretation of rise and fall times may depend on this threshold. Can the authors' test justify their threshold or report the sensitivity of their conclusions to alternative thresholds?
As the authors have a strong history of using simple models to study the energetics of gait, a (possibly speculative) discussion on how individual differences in specific features of locomotion (stabilization, leg swing, etc…) may have on the estimated valuations of time may improve readers' evaluation of the immediate generalizability of the paper's conclusions.
It would be interesting to see the following results, probably in a supplemental:
1. The individual values of cT for human participants and the corresponding preferred speeds.
2. Subjectspecific trajectories from Figure 4. Do any participants exhibit unexpected behavior?
https://doi.org/10.7554/eLife.81939.sa1Author response
Essential revisions:
1) The title asserts that "humans dynamically optimize walking speed to save energy and time". Directly substantiating this claim would require independently manipulating the (purported) energy and time cost of walking for human subjects, but these manipulations are not undertaken in the present study. What the Results actually report are two findings: a. (simulation) minimizing a linear combination of energy and time in an optimal control problem involving an invertedpendulum model of walking bouts that (i) start and stop at rest and (ii) walk at constant speed yields a gentlyrounded speedvstime profile (Figure 2A); b. (experiment) human subject walking bouts that started and stopped at rest had selfsimilar speedvstime profiles at several bout lengths after normalizing by the average duration and peak speed of each subject's bouts (Figure 4B). If the paper established a strong connection between (a.) and (b.), e.g. if speedvstime trajectories from the simulation predicted experimental results significantly better than other plausible models (such as the 'steady minC’T' an‘ 'steady accel' models whose trajectories are shown in Figure 2A), this finding could be regarded as providing indirect evidence in support of the claim in the paper's Title. Accordingly, it would be more accurate to assert 'brief human walking bouts look like trajectories of an invertedpendulum model that minimize a linear combination of energy and time' (although this is obviously too wordy). But unfortunately, the connection between (a.) and (b.) is only discussed qualitatively, and the other plausible models introduced in the Results are not revisited in the Discussion. The representative 'steady minC’T' trace in Figure 2A may be a real contender with the 'EnergyTime' trace for explaining the experimental results in Figure 4, but this candidate is rejected at the end of the thirdtolast paragraph in the "Model Predictions" subsection of Results based on the vague rationale that is never revisited. Reexamining and discussing these points will be important in a subsequent revision.
We agree with the comments about title, and have changed the title to “Optimization of energy and time determines dynamic speeds for human walking”. We believe this is accurate and consistent with the reviewer’s concerns.
We have also clarified why we reject the minCOT hypothesis. In Experimental Results, “In contrast to the minCOT hypothesis, the human peak speeds increased with distance, many well below the minCOT speed of about 1.25 m/s.” (The minCOT hypothesis always predicts only one speed.) And in Discussion, “Shorter bouts… exhibited a relatively brief peak slower than the typical minCOT speed… All such bouts also spent substantial time and energy in acceleration and deceleration…This contrasts with the minCOT hypothesis, which predicts only a single steady speed, and cannot explain how to start or end a bout.”
2) An additional limitation of the approach not discussed in the manuscript is that a fixed step length was prescribed in the simulations. Th“ "Optimal control formulation" subsection in "Methods" summarizes the results of a sensitivity analysis conducted by varying the fixed step length, but all results reported here impose a constantsteplength constraint on the optimal control problem. Although this is a reasonable modeling simplification for steadystate walking, it is less wellmotivated for the walking bouts considered here that start and stop at rest. For instance, the representative trial from a human subject in Figure 8 clearly shows initiation and termination steps that differ in length from the intermediate steps (visually discernable via the slope of the dashed line interpolating the black dots). Presumably different trajectories would be produced by the model if the constantsteplength constraint were removed. It is unclear whether this change would significantly alter predictions from either the 'EnergyTime' o‘ 'steady minC’T' model candidates. While the reviewers agreed that this change would entail substantial work that may be out of the scope of the present paper, they did agree that it is important to at least discuss this limitation.
We were not sufficiently we had also included step lengths increasing according to the preferred human step length vs. speed relationship. We have made several changes to clarify, in Model Predictions, Figure 3, optimal formulation in Methods, and Discussion. The specific changes are listed below (emphasis is used here only, and is not in actual text).
a. We have added “at humanlike step length” to the plainlanguage formulation.
b. In Model Predictions we have added, “Step length was nominally kept fixed, and then varied in parameter sensitivity studies below.”
c. In Figure 3 caption, “In the main plot, multiple predictions for different time valuations and step lengths are scaled and superimposed on each other… Three different step lengths including shorter and longer steps than nominal, and human preferred step length relationship (dashed lines); main plot also includes nominal step length.”
d. In the parameter sensitivity paragraph, “We considered step lengths s fixed at nominal (0.68 m), at slightly shorter and longer lengths (0.59 m and 0.78 m), and increasing with speed according to the human preferred step length relationship (see Methods for details).
e. In Methods, we moved the description of step length from the end of the subsection to earlier, and added a passage to clarify our hypothesis: “Step lengths were examined with a nominal fixed step length of 0.68 m, and sensitivity analyses performed with fixed lengths of 0.59 m and 0.78 m, and varying lengths following the human preferred step length relationship s=v^0.42 (Grieve, 1968). We have previously proposed that steptostep transitions account for most of the metabolic cost of human walking (Donelan et al., 2002), and that forced swing leg motions to modulate step length also contribute a nonnegligible, but lesser cost (Kuo, 2001) that is neglected here. This is not to dismiss this and other costs of locomotion, but merely to hypothesize that steptostep transitions will still dominate in transient walking bouts. The failure to include other costs, if sufficiently critical, should cause the model to make poor predictions.”
f. In Discussion, we have clarified that the mention of unmodeled forced leg motion is related to “step length modulation”.
3) Related to the above, the reviewer thought that the last two sentences in the abstract are too strong and should be rephrased to be more accurate and defensible.
We have reworded as follows (see emphasis):
“Individualdependent vigor may be characterized by the energy one is willing to spend to save a unit of time, which explains why some may walk faster than others, but everyone may have similarshaped trajectories due to similar walking dynamics. Tradeoffs between energy and time costs can predict transient, steady, and vigorrelated aspects of walking.”
4) Please explain the rationale for the competing objective functions (constant acceleration, in particular). The minCoT objective appears supported by literature, but it was clearly less plausible for transient gait than the energytime objective. The constant acceleration and minCoT objectives did not seem like terribly challenging competing hypotheses to refute. Perhaps the authors can test or comment on objectives that produce realistic families of speed profiles to the energytime hypothesis.
We have added more description of the steady acceleration hypothesis to Methods:
“The second alternative was a steady acceleration hypothesis, which contrasts with minC’T's immediate high acceleration. Here the acceleration is made as gradual as possible, albeit at the expense of a higher peak speed needed to travel the same distance and duration. Both of these alternatives help illustrate how different speed trajectories requires differing amounts of mechanical work, to be compared against the work produced by the EnergyTime hypothesis.”
We have added a Discussion paragraph about alternative models:
“It is instructive to consider what other models might explain or predict our experimental results. We did not explore more complex models here, but would expect similar predictions to result from any model based on pendulumlike walking and steptostep transitions. This includes those cited in the previous paragraph, as well as a family of such models including threedimensional motion…knees and biarticular actuators…, and plantarflexing ankles… It is also possible that more complex, musculoskeletal models also perform substantial work and expend energy for steptostep transitions, and might therefore agree with the present model. But here, relatively simple principles account for a fairly wide array of predictions (Figure 3), which are unlikely to result from happenstance. We are also unaware of any current hypothesis that could plausibly substitute for the present one. We therefore doubt if alternative models not based on pendulumlike principles could predict or reproduce these results, except with numerous fitted parameters.”
5) The rationale for defining rise and fall times based on 90% of the peak speed is unclear. The experimental acceleration profiles appear slightly steeper than the deceleration profiles, up to ~50% of the max speed. Consequently, the interpretation of rise and fall times may depend on this threshold. Can the authors justify their threshold or report the sensitivity of their conclusions to alternative thresholds?
There is nothing fundamental about the definition of rise/fall times. To clarify that these measures are only used descriptively, we have amended the Results to say “This was described by rise time…” which recapitulate the Methods “We also described walking durations in terms of rise and fall times” (emphasis is not in the actual text). The definition here is drawn from the convention in control systems (10% and 90%, https://en.wikipedia.org/wiki/Rise_time), but other definitions could also be appropriate.
We made no strong claims based on these measures, which were defined mainly to provide some visualization (Figure 5B). The only statements were (in Results), “As a fraction of each bout's duration, the rise and fall times appeared to take up a greater proportion for shorter bouts, and only a very small proportion was spent at steady speed.
Conversely, cruise time took up a greater proportion of the time for longer bouts.” The “appeared to” was intentional, to indicate a qualitative observation based on nonfundamental rise/fall times. The second sentence is a continuation of the first, and so we did not bother add another qualifier there. We feel this is a defensible statement, because all trials from all subjects are shown in Figure 4, and we do not think the observation is highly sensitive to the definition of rise/fall time.
6) As the authors have a strong history of using simple models to study the energetics of gait, a (possibly speculative) discussion on how individual differences in specific features of locomotion (stabilization, leg swing, etc…) may have on the estimated valuations of time may improve readers' evaluation of the immediate generalizability of the paper's conclusions.
We have expanded the speculative Discussion of energetics and stability:
“The simple walking model also only includes a crude representation of steptostep transitions, which we have crudely estimated to account for as much as 70% of net metabolic cost in steady walking (Donelan et al., 2002). We did not include other factors such as forced leg motion and step length modulation (up to say 33% of cost, Doke et al. 2005) that likely also affect energetic cost, and could therefore be used to test the valuation of time. Nor did we include factors such stability (Bauby and Kuo, 2000; Donelan et al., 2004, Rebula et al., 2017) and threedimensional motion Donelan et al., 20010, despite being part of our previous models, because we believe them to contribute little to the present task. In fact, because the optimal control model successfully completes the walking task, its feedforward motion attains implicit stability, which could reduce (but not eliminate) the need for feedback stabilization (Darici and Kuo, 2022 a, b).”
7) In accordance with eLife policies, the authors need to release their data and simulation code, as well as provide details about their experimental apparatus.
Yes, we intend to share both data and optimization code if manuscript is accepted, as indicated in Data Availability statement. This information is under preparation.
8) The manuscript would benefit from careful proofreading prior to publication  the reviewers caught a number of typographical and grammatical errors.
We apologize for the errors and have made an effort to fix them.
Reviewer #1 (Recommendations for the authors):
Almost all of the following has been addressed above in Editor comments. Here we respond to a few comments with additional information.
Conclusion
– It might be fun to cite the (purported) origin of the "time is money" aphorism.
https://en.wikipedia.org/wiki/Time_is_money_(aphorism)
The origin of the aphorism is interesting, but we prefer not to include it.
Methods
– I was absolutely delighted to learn about JuMP! Thank you for teaching me – I commend your use of opensource tools and hope you'll make your source code available to the community.
We do intend to share the code and the data. The entire toolchain for the optimization is open source.
Reviewer #2 (Recommendations for the authors):
All of this has been addressed above in Essential comments.
https://doi.org/10.7554/eLife.81939.sa2Article and author information
Author details
Funding
Natural Sciences and Engineering Research Council of Canada (CRC Chair)
 Arthur D Kuo
Université Lille 1  Sciences et Technologies (Tier 1)
 Arthur D Kuo
Natural Sciences and Engineering Research Council of Canada (Discovery award)
 Arthur D Kuo
Dr. Benno Nigg Research Chair
 Arthur D Kuo
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Senior Editor
 Joshua I Gold, University of Pennsylvania, United States
Reviewing Editor
 Gordan J Berman, Emory University, United States
Reviewer
 Sam Burden, University of Washington, United States
Version history
 Received: July 17, 2022
 Preprint posted: July 18, 2022 (view preprint)
 Accepted: January 11, 2023
 Accepted Manuscript published: February 13, 2023 (version 1)
 Version of Record published: March 21, 2023 (version 2)
Copyright
© 2023, Carlisle and Kuo
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics

 1,122
 Page views

 170
 Downloads

 3
 Citations
Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading

 Computational and Systems Biology
Cellfree DNA (cfDNA) tests use small amounts of DNA in the bloodstream as biomarkers. While it is thought that cfDNA is largely released by dying cells, the proportion of dying cells' DNA that reaches the bloodstream is unknown. Here, we integrate estimates of cellular turnover rates to calculate the expected amount of cfDNA. By comparing this to the actual amount of cell typespecific cfDNA, we estimate the proportion of DNA reaching plasma as cfDNA. We demonstrate that <10% of the DNA from dying cells is detectable in plasma, and the ratios of measured to expected cfDNA levels vary a thousandfold among cell types, often reaching well below 0.1%. The analysis suggests that local clearance, presumably via phagocytosis, takes up most of the dying cells' DNA. Insights into the underlying mechanism may help to understand the physiological significance of cfDNA and improve the sensitivity of liquid biopsies.

 Computational and Systems Biology
 Evolutionary Biology
Synthetic autotrophy is a promising avenue to sustainable bioproduction from CO_{2}. Here, we use iterative laboratory evolution to generate several distinct autotrophic strains. Utilising this genetic diversity, we identify that just three mutations are sufficient for Escherichia coli to grow autotrophically, when introduced alongside nonnative energy (formate dehydrogenase) and carbonfixing (RuBisCO, phosphoribulokinase, carbonic anhydrase) modules. The mutated genes are involved in glycolysis (pgi), centralcarbon regulation (crp), and RNA transcription (rpoB). The pgi mutation reduces the enzyme’s activity, thereby stabilising the carbonfixing cycle by capping a major branching flux. For the other two mutations, we observe downregulation of several metabolic pathways and increased expression of native genes associated with the carbonfixing module (rpiB) and the energy module (fdoGH), as well as an increased ratio of NADH/NAD^{+}  the cycle’s electrondonor. This study demonstrates the malleability of metabolism and its capacity to switch trophic modes using only a small number of genetic changes and could facilitate transforming other heterotrophic organisms into autotrophs.