Simulating locomotion under anatomical and mechanical constraints

Characterizing the locomotor behavior of animals is essential for any study of their biology and ecology. Locomotor behavior comprises how an animal navigates its environment, interacts with the substrate, and moves its body and appendages in a controlled and coordinated manner. Robust locomotion requires coordinated movement under anatomical and mechanical constraints. It also comprises the ability to adapt quickly to changing environmental conditions. This includes gait modulation (typically in response to small disturbances or inhomogeneities in the substrate), as well as gait switching (e.g., in response to a threat). Insects are exceptionally robust walkers, and they exhibit a wide range of anatomical and functional adaptations to the environmental constraints posed by their natural habitats. Thus, they serve as excellent models for studying the evolutionary interplay between body morphology and locomotion under mechanical constraints. In this work, we demonstrate the usefulness of computer models for the study of animal locomotion. We focused on two insects that reside in distinct habitats and, accordingly, have developed different lifestyles and morphological adaptations. The well-studied desert locust walks and hops on various surfaces, and flies, typically in swarms, whereas the less-studied mole cricket lives solitarily underground and digs tunnels. Consequently, the mole cricket features significant anatomical adaptations, including shovel-shaped forelegs for digging, large thrust-producing hind legs, a cylindrical, sclerotized body, and a pointed head. To capture the robustness of adaptive locomotion in simulation, we study gait transition and modulation in insects and employ reinforcement learning to train agents for multiple gaits, focusing on the strategies used to execute these transitions and modulations, as well as the external stimuli that trigger them. Overall, the implications of this study are two-fold. First, our comparative approach enables us to address form-function knowledge gaps and study the mechanisms underlying robust and adaptive locomotion. Second, training policies using anatomically and mechanically grounded models in a physical simulation environment can significantly reduce the training time for robots in real-world settings (i.e., sim-to-real transfer).