This study was designed to understand higher-level processes in balancing skills, evident in the decision-like switching between a finite set of manual command alternatives for balancing oneself or an external object (Loram and Lakie 2002; Cabrera and Milton 2004; Bottaro et al. 2008; Panic et al. 2015; Zgonnikov and Markkula 2019; Vimal et al. 2020; Wang et al. 2022). Our approach included (1) the use of the VIP task, which depends little on musculoskeletal biomechanics, reflexive stiffness and coordination, and vestibular/proprioceptive acuity, but more heavily expresses higher-level factors in sensorimotor skill, (2) a novel conceptual framework that emphasizes the spatiotemporal regimes where command timing and accuracy are critical, and (3) comparisons across age groups where balancing performance is known to decline in a manner that depends on both peripheral factors (Nardone et al. 1995; Richardson et al. 2014; Jahn 2019) and central sensorimotor integration and decision-making (Muller et al. 2004; Li et al. 2018).

Effects of age, joystick gain, and delay across whole trials. Our finding that YA outperform OA on whole-trial measures of VIP balancing is consistent with the results of comparable tests (Jagacinski et al. 1995; Liao et al. 1997). YA fell (contacted the fall boundaries) less often and oscillated with lower rates. In addition, YA fell more softly (with a lower average velocity ≈ 77 ± 18°/s) than OA (≈ 108 ± 33°/s), p < 0.001, W = 617, Wilcoxon rank-sum test. Our significant main effects of both gain and delay are consistent with previous conclusions that reflex force and latency are risk factors in self-balancing and visuo-manual tracking (Jagacinski et al. 1995; Nardone et al. 1995; Liao et al. 1997; Richardson et al. 2014; Jahn 2019). As expected, our artificially long delays increased the fall rate and the variance of VIP sway magnitude and rate. We also expected performance to be degraded by joystick gain increases because extensive parametric testing of the VIP paradigm had shown that the lower of our two gain conditions was optimal while the hyper gain elicited operator-induced oscillations. The artificially high joystick gains increased the fall rate and the variance of the VIP sway rate, paralleling the expectation of muscle strength being a factor.

In addition, the full pattern of joystick gain and delay effects of our experimental manipulation of VIP extend past studies, as they also rule out a purely peripheral explanation of performance. The maximum delay we imposed – 60 ms – is longer than any published estimate of age-related intrinsic visuomotor delay, so a purely peripheral explanation would predict worse performance in YA with a 60 ms delay than in OA with 0 experimental delay plus less than 60 ms intrinsic delay, especially under hyper gain conditions. However, YA with 60 ms joystick delays fell less than OA with no delay and comparable joystick gain (Fig. 2, top left). Our whole-trial results also extend the literature by showing that D commands, which have not been reported in previous object balancing studies, are present for both age groups, with a prevalence of about 5% in YA and 10% in OA (averaged across whole trials). Furthermore, I, CR, and D commands are unaffected by joystick delay but are reciprocally affected by age: OA > YA for I and D, and OA < YA for CR. By contrast, A commands are unaffected by age but increase with delay. The fact that A commands are not affected by age implies OA and YA do not differ in capacity for predicting pendular behavior. This separability of factors affecting I, CR, and D vis-à-vis A in whole-trial analysis, combined with the fact that I, CR, and D are definable in Fall quadrants but A only in Safe quadrants, suggest that balancing dynamics—task-relevant co-evolution of pendulum (plant output) and joystick (controller) states—are factors in balancing performance.

Age differences in control maneuver types depend on dynamic regimes of balancing. The absence of interaction effects among age, joystick gain, and delay factors on command types in whole-trial univariate analyses led us to combine different gain and delay conditions for each balancing regime. Additionally, we reasoned that command switching would be governed by time pressure, which would increase when falls are imminent. For the task difficulty (KP) we employed, the whole-trial rate of switching between the four functional command types was roughly 2 Hz, manifested as the number of color-coded command sub-sequences in the typical 30 s trial plotted in Fig. 1a, without the transitions caused by the pendulum crossing a quadrant. We partitioned the time series of pendulum balancing into three distinct dynamic regimes: Safe, Saved, and Failed, which are defined by time pressure and command selection. In the Safe regime, the control could be either I, CR, or A; in the Saved and Failed regimes, control could involve I, CR, or D maneuvers.

The overall evolution of all command types for both age groups exhibited monotonic transitions as a function of the angular deviation of the pendulum from the DOB. Figure 3 illustrates that for balancing in all regimes, the percentage of CR commands transitioned smoothly from a minimum near the DOB to a maximum when displaced 60° (at the fall boundaryFootnote 4). The percentage of I and D commands evolved in reverse, from high to low. In other words, consistent patterns of command switching are observed with increases with deviation from the upright goal, which has never been reported before, and it occurs for both age groups. Furthermore, age-related differences in the evolution of CR commands exist only in Failed balancing. YA made significantly more and earlier CR than OA only in the Failed balancing regime. Near the boundary, YA also made significantly fewer D movements and were less inactive than OA only in the Failed regime. These age group differences in novel patterns of command switching within the Failed balancing regime parallel age differences in falling; the group with fewer CR and more D commands (only in the Failed balancing regime) falls more often (falls occur only in Failed balancing). We found no age differences in the speeds at which the age groups enter the Fall quadrant, meaning that age differences evolve within the Failed balancing regime.

Age-dependent, regime-specific, and time-critical switching between balancing command alternatives. The observation that the age differences in command switching as a function of pendulum position occurred only in the dynamic regime where falls happen, and were reminiscent of logistic functions, motivated post-hoc analyses probing whether switching is time-critical. Guided by psychophysical and neurophysiological switching patterns for saccadic decisions (Schall and Thompson 1999; Salinas et al. 2010; Costello et al. 2013), we assessed whether VIP command switching improves or declines as the time before a fall shortens, leading to less time available for decision-making. This would predict that, in Failed balancing, (1) the rate of inactivity (I) should decrease (because something must be done) as the pendulum gets closer in time to contacting a fall boundary, and after the threshold for action is exceeded, (2) YA, who perform better, should switch faster and more successfully (increase CR and decrease D) than OA. We computed command probability as a function of time left to fall (TLTF), assessed logistic curve fits, and compared the fits with functionally relevant summary statistics. Logistic functions were the best fits (ranked 1st among 9 plausible monotonic functions) for I and CR for both age groups, and for D, it was the best for YA and the second-best fit for OA.

Overall, both post-hoc hypotheses about Failed balancing were supported. The time of logistic transition from a low to high proportion of fall-saving CR commands (\(_^\)) was temporally closer to a fall (later) in OA than YA; CR commands integrated over the entire range of TLTF (\(_^\)) were less prevalent in OA than YA who fell less; immediately before contacting a fall boundary, OA made a smaller percentage of CR commands (\(_^\)), and a larger percentage of fall-promoting D commands (\(_^\)) and fall-capitulating I commands (\(_^\)) than YA. Both groups’ command proportions were statistically identical when they had abundant TLTF upon entering the Fall quadrant with identical velocity.

The post-hoc hypothesis of age-dependent, time-critical command switching makes two supplementary predictions: 1) within Fall quadrants, command switching should be better and faster in Saved balancing, where falling is averted than in Failed balancing, which always ends in a fall, despite identical initial pendulum and command states, and 2) in Saved balancing where both groups escape the danger of falling there should be no differences in command switching. The reverse of command percentage as a function of displacement (see Fig. 3) is a rough proxy for percentage as a function of TLTF, and also, it can be calculated for all balancing regimes. Therefore, the regimes can be directly compared in terms of \(_^\), \(_^\), and \(_^\). Our analyses, supported both supplementary predictions. For the first question, all three CR variables differed between the Failed and Saved regimes in fall-saving directions, but Saved and Safe did not differ. For the second question, all three CR variables were significantly more fall-saving in YA than OA in Failed balancing (p ≤ 0.004 at least), but there were no such age differences in Safe and Saved balancing.

Figure 5a summarizes the observed coordinated increase in the percentage of fall-saving CR commands and decrease in fall-promoting D commands as TLTF approaches 0. The ratio %CR/(%CR + %D) on the y-axis represents the proportion of fall-saving commands relative to all active commands leading up to a fall. This ratio gets more favorable as falls become more imminent, more dramatically for YA than OA. Including inactivity in the denominator would rescale both curves without altering the age differences because %I as a function of TLTF does not differ by age group, except at the ultimate moment before a fall (\(_^\)), where it would magnify the age differences. This illustrates that OA are as competent as YA in switching between action and inaction, but OA become more error-prone in choosing which and how much action to take, and they do so later as TLTF approaches zero. Figure 5b plots the area under composite %CR/(%CR + %D) curves of 5a for each individual on the y-axis as a measure of overall fall-saving behavior against the propensity to fall. The negative slope proves an association between better Failed regime command structure (more CR and less D) and better task-level VIP performance (fewer falls). The statistically indistinguishable linear fits for OA (gray circular dots) and YA (black squares) indicate the same relationship holds for both age groups. The distinct cluster of points in the lower right part of Fig. 5b suggests that a subset of “high fallers” may exist—mostly OA but at least one YA—who show extreme dysfunctional coordination of CR and D commands relative to the general population.

a The ratio of fall-saving CR commands to all active commands as a function of TLTF in Failed balancing, averaged for OA and YA. b Scatterplots of the area under the curve of the plots in panel a (normalized to the highest value in the entire data set) versus the average fall rates per participant, for OA (gray circles) and YA (black squares). Neither the slopes nor y-intercepts of the linear fits differ by age

Implications of age-dependent, regime-specific, time-critical switching between balancing command alternatives. OA and YA do not differ in task-level performance or command dynamics in Safe and Saved balancing, they differ only in Failed balancing. In the Safe quadrants, falls are not an immediate threat, and both OA and YA use identical command switching to stay equivalently well around the upright DOB. In the Fall quadrants, the initial pendulum and joystick conditions are the same for both OA and YA until non-optimal CR and D commands lead to an age-dependent divergence of Saved and Failed balancing. Any mechanistic explanation for the divergence must account for OA and YA being identical in Safe balancing, identical in how they escape imminent falls (Saved balancing) but different in how they succumb to imminent falls (Failed balancing). For example, delayed or less precise encoding of the pendulum state—it’s angular position and velocity—might be evoked to explain the OA deficit in Failed balancing but is incompatible with the statistically similar performance of OA and YA in Safe and Saved regimes. In addition, such regime-specific age differences in switching are improbable if under the sole control of linear optimal feedback mechanisms incorporating delays (Milton et al. 2009a), drift and catch (Milton et al. 2009b). If the parameters of such models were set to reflect age differences in peripheral biomechanical and neurophysiological characteristics, then they should result in age differences in performance independent of the balancing regime instead of the regime-dependent age difference we found. These novel findings complement current notions of human intermittent control, and the key reason for these advances is our novel method of functionally characterizing the phase space rather than treating it as uniform. Furthermore, our demonstrations of age- and regime-specific behavior does not exclude the possibility that models without regime-specific parameterization also operate concurrently.

Our discovery of the age-dependent logistic structure of I and CR or D command dynamics in Failed balancing is consistent with a decision-like process. A lack of significant difference between OA and YA when there is no imminent threat of falling (Saved and Safe regimes) implies that balancing may operate in dual decision modes: “oscillate about the DOB” when TLTF is not pressing enough to evoke action, and “fall prevention mode” when under time pressure. In both cases, there could be a logistic, probabilistic decision to act or not, and after the decision to act advances far enough, a choice must be made between CR and D in the Fall quadrants or between CR and A in the Safe quadrants. Such dual, simultaneous modes have been proposed to explain the performance and joystick commands in a MARS self-balancing experiment made difficult by depriving participants of information about self-tilt relative to gravity (Vimal et al. 2019). Zgonnikov and Markkula (2019) previously concluded that decision-making processes that result in transitions from inactivity to corrective action operate when manually controlling a mechanical inverted pendulum. However, they did not model decisions between corrective (CR) versus destabilizing (D) actions in the Fall quadrants or versus anticipatory (A) actions in the Safe quadrants. We note in this context that we have chosen to discriminate recurrent, discrete periods of command “activity/inactivity” without using the term “intermittency,” which often connotes clock-like or phase-dependent switching. In addition, it bears repeating that D commands have not been documented outside the present study and previous MARS studies from our laboratory. It is beyond the scope of this paper to schematize all possible decision transitions, but inspection of Fig. 1 shows that transitions can occur bidirectionally between inaction and action, as well as between the three types of actions (CR, A, and D).

Relationship of VIP results to bipedal balance and falling, and limitations. The VIP task is only a partial analog to bipedal balance, and the costs of VIP “falling” are much lower than the cost of real falling. However, the present results provide new insights for investigating falling in at-risk populations. We are unaware of any previous suggestion that falling may be due to frankly destabilizing muscle activation, as opposed to the correct direction being executed too weakly, too late, and/or too intermittently. The VIP paradigm spotlights the continuous, high rate of command switching needed, as well as the high relevance of imminent falling. We are currently examining the possibility that destabilizing commands in the Fall quadrants are confusions with recently activated decision processes in the Safe quadrants, in the VIP task, the MARS self-balancing task, and the real bipedal stance. Such confusions are plausible because they appear in high-rate serial saccade and keyboarding tasks (Rosenbaum et al. 1986; Emeric et al. 2007). Our simple, single-pivot VIP paradigm results have potential parallels to command errors and command switching in bipedal stance. Medial–lateral sway is physically governed by the weight distribution between the two feet and the pressure distribution under each foot (Winter et al. 1996; Rougier 2007), among other factors. Even when one leg gradually assumes the majority of body weight support, it almost instantly engages in contributing a biomechanically disproportionate dominance of the net dynamic center of pressure (while the other leg disengages, acting only as a vertical strut) (Bakshi et al. 2019, 2020). Such concurrent, coordinated command switching is not prominent in the raw center of foot pressure traces but can be decoupled with appropriate two-foot force measurements, and the evolution of different leg muscle command types can be analyzed as a function of the balance regimes defined for the VIP.

In a different vein, the ample literature documenting dual-task, cognitive-postural deficits could be explained by competition between central decision-like balancing mechanisms and dual-task central processing (Muller et al. 2004), but the peripheral postural process would not be expected to compete with secondary cognitive tasks. Finally, it is possible that age differences in command transitions with diminishing TLTF may reflect subjective estimates of risk or fear of falling in the VIP task, which are factors in bipedal fall risk. Our OA participants had less exposure to joysticks than YA participants, but there were no significant differences in fear of falling. In future studies, we are interested in distinguishing the decision-making differences between high fallers, such as those in the lower right of Fig. 5b, and proficient performers, independent of age, in the VIP task, MARS self-balancing, and bipedal stance.