Effects of Mass Distribution and Center-of-Mass Position on Reaching Gait in a Parasaurolophus Locomotion Simulation Using Model-Free Deep Reinforcement Learning

Abstract

Soft tissues do not fossilize, so dinosaur mass reconstructions are not unique. Previous static reconstructions have examined how alternative body-volume assumptions affect total mass and center-of-mass (CoM) position. Dynamic locomotion simulations, in contrast, have evaluated generated gaits and speeds. The effects of mass distribution on gait have therefore not been directly compared under matched task and training conditions. This study aimed to build a simulation environment that can generate and analyse dinosaur gaits using model-free deep reinforcement learning (a method that learns actions through trial and error guided by reward, without prespecified rules of movement) and to use it to compare the effects of mass distribution (the allocation of mass among body segments) on reaching gait (the gait learned to reach a target). Here, a model-free deep reinforcement learning locomotion simulation was used to vary only the density allocation of head-neck and tail segments while holding the skeletal model, reward function, control-system configuration, and training conditions constant. Gaits were evaluated by CoM at the reaching pose, forelimb contact rate, and bipedal stance rate. Within the Both mass-allocation path, more cranial CoM positions were associated with higher forelimb contact rates, whereas more caudal CoM positions did not make bipedal stance obligatory. Moreover, changing which segment group received the density adjustment altered both CoM at the reaching pose and gait even at the same target CoM. These findings show that, within this model and learning configuration, target CoM alone is insufficient to determine gait uniquely.