NeuroMechFly v2: simulating embodied sensorimotor control in adult Drosophila

Area of Science:

• Neuroscience
• Computational Biology
• Robotics

Background:

• Neuromechanical models are crucial for understanding animal behavior, but existing models often focus narrowly on motor control.
• Hierarchical sensorimotor control, involving brain-motor system interaction, remains less explored in computational models.
• The fruit fly (Drosophila) serves as a powerful model organism for studying complex behaviors.

Purpose of the Study:

• To expand the capabilities of the NeuroMechFly neuromechanical modeling platform.
• To enable the simulation of sensory inputs (vision, olfaction) and complex environmental interactions (terrain, leg adhesion).
• To facilitate the development of biologically inspired and machine learning-based controllers for autonomous agents.

Main Methods:

• Development of NeuroMechFly v2, incorporating vision, olfaction, ascending motor feedback, and complex terrain navigation with leg adhesion.
• Construction of biologically inspired controllers for path integration and head stabilization using ascending feedback.
• Application of reinforcement learning to train a controller for multimodal navigation tasks.
• Bio-realistic modeling of complex odor plume navigation and fly-fly following using a connectome-constrained visual network.

Main Results:

• NeuroMechFly v2 successfully integrates multiple sensory modalities and complex environmental factors into neuromechanical simulations.
• Biologically inspired controllers demonstrated effective path integration and head stabilization.
• Reinforcement learning enabled a controller to perform a multimodal navigation task.
• Advanced simulations showcased realistic odor plume navigation and social interaction behaviors.

Conclusions:

• NeuroMechFly v2 significantly advances the potential of neuromechanical modeling for neuroscience research.
• The platform accelerates the discovery of explanatory models of nervous system function.
• It provides a foundation for developing sophisticated machine learning controllers for robotics and artificial agents.