MotorNet, a Python toolbox for controlling differentiable biomechanical effectors with artificial neural networks.

Area of Science:

• Computational Neuroscience
• Robotics
• Machine Learning

Background:

• Artificial neural networks (ANNs) are crucial for understanding brain function but face limitations in neural control of movement research.
• Current methods require separate platforms for ANNs and biomechanical simulations, and non-differentiable effectors restrict training algorithms.

Purpose of the Study:

• To develop an integrated, user-friendly toolbox for training ANNs on differentiable, biomechanically realistic models for motor control tasks.
• To address the impracticalities of existing methods by providing a unified and efficient framework.

Main Methods:

• Developed MotorNet, an open-source Python toolbox utilizing PyTorch for creating complex, differentiable, and biomechanically realistic effectors.
• Designed with ease of installation, a high-level API, and a modular architecture for flexibility.
• Tested on standard motor control models, demonstrating rapid training on typical hardware.

Main Results:

• MotorNet enables the creation of differentiable biomechanical effectors, facilitating broader application of ANN training methods.
• The toolbox allows for rapid training of ANNs on motor control tasks, significantly reducing setup and computation time.
• It integrates seamlessly with the PyTorch framework, benefiting from AI advancements.

Conclusions:

• MotorNet overcomes key limitations in computational neuroscience for motor control research by providing a unified, efficient, and accessible framework.
• The open-source nature encourages community contributions, accelerating innovation in neural control and AI-driven biomechanics.
• Facilitates research by reducing overhead for new and established computational teams, allowing focus on conceptual advancements.