Control strategies for rapid, visually guided adjustments of the foot during continuous walking

Abstract

When walking over stable, complex terrain, visual information about an upcoming foothold is primarily utilized during the preceding step to organize a nearly ballistic forward movement of the body. However, it is often necessary to respond to changes in the position of an intended foothold that occur around step initiation. Although humans are capable of rapidly adjusting foot trajectory mid-swing in response to a perturbation of target position, such movements may disrupt the efficiency and stability of the gait cycle. In the present study, we consider whether walkers sometimes adopt alternative strategies for responding to perturbations that interfere less with ongoing forward locomotion. Subjects walked along a path of irregularly spaced stepping targets projected onto the ground, while their movements were recorded by a full-body motion-capture system. On a subset of trials, the location of one target was perturbed in either a medial–lateral or anterior–posterior direction. We found that subjects were best able to respond to perturbations that occurred during the latter half of the preceding step and that responses to perturbations that occurred during a step were less successful than previously reported in studies using a single-step paradigm. We also found that, when possible, subjects adjusted the ballistic movement of their center of mass in response to perturbations. We conclude that, during continuous walking, strategies for responding to perturbations that rely on reach-like movements of the foot may be less effective than previously assumed. For perturbations that are detected around step initiation, walkers prefer to adapt by tailoring the global, pendular mechanics of the body.

Appendix: Trajectory detection with MANOVA

To avoid Type-I errors when performing the search for deviations along the acceleration trajectories, a nominal value for expected “hit rates” was needed. That is, a threshold was needed that would quantify how many significant deviations in a row would be detected in the trajectory data by chance. To find this threshold, we employed a method developed by Dale et al. (2007) which used a bootstrap search for the nominal runs of significant deviations in a random set of data with the same physical and statistical properties.

To start, trajectories were time-normalized using a spline interpolation, such that the previous-step, current-step, and next-step portions were each 100 datapoints long. Thus, each point along the trajectory corresponded to a percentage of the step and could be directly compared. For each subject, average trajectories for each of the 17 experimental conditions (four perturbation times, four perturbation directions, and one unperturbed condition) were computed by taking the mean across each timepoint in all trajectories for each condition. This led to 12 subjects × (4 timing × 4 direction + 1 control) = 204 trajectories. These 204 trajectories were used to construct a series of simulated “experiments”.

Each simulated experiment involved comparing one timing and direction condition pair to control trajectories (e.g., perturbations at 0.99-steps in the medial direction compared to unperturbed walking). This entailed 24 trajectories: one manipulated and one controlled per subject. For each timepoint along the trajectories, the subject scores for the manipulated and control were used to create hypothetical distributions using the mean and standard deviations across subjects from which new simulated subjects could be drawn. For example, at each timepoint, a group of 12 x, y, and z acceleration values were drawn for the perturbed and unperturbed trajectories from normal distributions defined by the respective means and standard deviations. This gave new acceleration values for the perturbed trial and unperturbed trial. A repeated-measures multivariate analysis of variance (MANOVA) was used to decide if there was a significant difference between the perturbed and unperturbed trajectories at this timepoint, where the x, y, and z accelerations were taken as joint outcome measures. The p-value of this test was logged, and the procedure moved to the next timepoint and repeated.

This process was repeated along the entire length to the perturbed and unperturbed trajectory, and a p-value at each timepoint was logged. This trajectory of p-values could then be analyzed to determine how many times a run of p-values with a specific length n occurred. For example, there might 10 runs of 1 p < 0.05 alone, 6 runs of 2 p < 0.05 in a row, 4 runs of 3 p < 0.05 in a row. The process was repeated for each combination of perturbed trial and control trial, producing a run count for each condition. This entire process was then repeated 10,000 times. This produced 10,000 p-value run counts for each condition in the experiment. The number of times a particular run of length n was observed out of 10,000 was then used to quantify the likelihood of observing a run of length n by chance in the collected data. Runs of four significant p-values in a row were found to occur less than 1% of the time, so four was chosen as the threshold by which a significant deviation between trajectories would be identified. Thus, the first point at which two trajectories were found to be significantly different for four or more time points in a row along the length of trajectories was determined to be an accurate point of divergence. Because the MANOVA used to compute this divergence point consumed all subject data, no variance metric was obtainable for this particular measure.