Humans can walk economically on uneven terrain without deliberative optimization with a Simple Feedback Control Policy
Humans proactively anticipate and adjust for uneven terrain as they traverse it. They modulate their forward speed and momentum ahead of upcoming steps, perhaps to optimize energy expenditure as they prefer to do for level ground. But actual optimization is a complex, deliberative process and seems not to describe how human act in real time during continuous walking. Here we show that human uneven walking strategies are predicted well by an control scheme that requires no deliberation yet is nearly optimal. The experimentally measured speed trajectory for a single upward step (a kernel) predicts the strategy for a complex terrain sequence, by repeatedly delaying and weighting (convolving) it for each step height. Conversely, the same strategy can be deconvolved to predict for six other terrains about as well as deliberative energy minimization Composability is predicted by a simple model of walking, which requires work to redirect the body center of mass velocity between pendulum-like steps. These dynamics obey linear superposition, so that a simple kernel, learnable through experience, is sufficient to compose other strategies with lookahead information about upcoming step heights. Walking dynamics may allow humans to act quickly and nearly optimally without thinking slowly.