Learning Inverse Dynamics of Multi-segmented Larval Crawling Neuromechanics
Weihao Tang, Han Zhang, Zhefeng Gong, Nenggan Zheng · 2024
Neuromechanical modeling that bridges neural activity and locomotion patterns is essential for studying locomotion control. The Drosophila larva, with its diverse softbodied locomotion, serves as an ideal subject for biomimetic research. Constructing a neuromechanical model of Drosophila larvae presents two primary challenges: efficiently simulating the larva’s soft, multi-segmented body and controlling locomotion within a large action space by actuating muscle groups. This work presents a three-dimensional neuromechanical model of the larva and introduces an inverse dynamics learning method for locomotion control. Our approach efficiently simulates the larva’s multi-segmented, deformable body by integrating computational modules for soft tissues, muscles, and friction. We employ a parameterized controller to reproduce peristaltic waves and use a hybrid artificial neural network architecture to learn inverse dynamics from partially observed states. Balancing biological interpretability with modeling complexity, this model advances the understanding of brain-body-environment interactions in Drosophila larvae, contributing to nature-inspired intelligent systems research.