AI-CPG: Adaptive Imitated Central Pattern Generators for Bipedal Locomotion Learned Through Reinforced Reflex Neural Networks

Striding bipedalism is a key derived behaviour of hominids that possibly originated soon after the divergence of the chimpanzee and human lineages. Although bipedal gaits include walking and running, running is generally considered to have played no major role in human evolution because humans, like apes, are poor sprinters compared to most quadrupeds. Here we assess how well humans perform at sustained long-distance running, and review the physiological and anatomical bases of endurance running capabilities in humans and other mammals. Judged by several criteria, humans perform remarkably well at endurance running, thanks to a diverse array of features, many of which leave traces in the skeleton. The fossil evidence of these features suggests that endurance running is a derived capability of the genus Homo, originating about 2 million years ago, and may have been instrumental in the evolution of the human body form.

While neuroscientists identify increasingly complex neural circuits that control animal and human gait, biomechanists find that locomotion requires little control if principles of legged mechanics are heeded that shape and exploit the dynamics of legged systems. Here, we show that muscle reflexes could be vital to link these two observations. We develop a model of human locomotion that is controlled by muscle reflexes which encode principles of legged mechanics. Equipped with this reflex control, we find this model to stabilize into a walking gait from its dynamic interplay with the ground, reproduce human walking dynamics and leg kinematics, tolerate ground disturbances, and adapt to slopes without parameter interventions. In addition, we find this model to predict some individual muscle activation patterns known from walking experiments. The results suggest not only that the interplay between mechanics and motor control is essential to human locomotion, but also that human motor output could for some muscles be dominated by neural circuits that encode principles of legged mechanics.

The transition from aquatic to terrestrial locomotion was a key development in vertebrate evolution. We present a spinal cord model and its implementation in an amphibious salamander robot that demonstrates how a primitive neural circuit for swimming can be extended by phylogenetically more recent limb oscillatory centers to explain the ability of salamanders to switch between swimming and walking. The model suggests neural mechanisms for modulation of velocity, direction, and type of gait that are relevant for all tetrapods. It predicts that limb oscillatory centers have lower intrinsic frequencies than body oscillatory centers, and we present biological data supporting this.