Abstract
Traditional theories of animal locomotion rely on Central Pattern Generators (CPGs)—hard-wired neural circuits that produce rhythmic activity. However, these models often lack the adaptability required to handle varying body mechanics or physical changes. This study demonstrates that a simple, randomly connected neural network can learn to generate complex rhythmic movements by interacting with a mechanical body. By employing the Bienenstock–Cooper–Munro (BCM) learning rule, the model transitions from chaotic motion to stable, coordinated oscillations. These findings suggest that locomotion may be an emergent property of sensorimotor interaction rather than a strictly genetic program.
Introduction
The prevailing neuroscience paradigm suggests that spinal cord circuits contain specialized "pacemaker" neurons responsible for rhythmic movement. While this explains how rhythms are maintained, it struggles to account for the adaptability seen in nature. Animals with similar genetic blueprints possess diverse body shapes, and many can adapt to significant mechanical alterations, such as tendon transfers.
The authors investigated whether a generic neural network could learn to generate these rhythms through physical interaction. They hypothesized that a homeostatic learning rule, originally used to describe the visual cortex, could organize a spinal-like network to produce stable motor outputs.
Methods
The researchers developed a closed-loop simulation environment consisting of a neural network coupled with mechanical models.
The Neural Network and BCM Rule
The network comprised eight non-spiking neurons with all-to-all connectivity (excluding self-connections). Initially, weights were assigned randomly, and no neurons possessed intrinsic pacemaker properties. Learning was driven by the Bienenstock–Cooper–Munro (BCM) rule. This rule encourages neurons to become selective by pushing their firing rates toward a bimodal distribution—either highly active or quiet—avoiding intermediate states.
Mechanical Systems
The network was tested on two distinct mechanical setups:
- Independent Pendulums: Two separate pendulums representing decoupled legs.
- Double Pendulum: A two-segment chain representing a multi-jointed limb.
Neurons received sensory feedback (joint angles and velocities) and generated motor commands (torque).
Training and Evaluation
The model underwent a 2,000-second training phase where it was stimulated with random motor "pokes" to trigger the BCM rule. Performance was evaluated by comparing 100 randomized test inputs before and after the learning period.
Key Results
The application of the BCM rule successfully transformed the system's behavior from decaying or chaotic motion into stable, rhythmic oscillations.
Emergent Coordination
A striking result was the spontaneous emergence of alternating patterns. In the independent pendulum test, the network learned to move the "legs" in a left-right alternating sequence despite the lack of a physical connection between them. The network discovered that mutual inhibition was a stable state that satisfied the BCM rule's homeostatic requirements.
Phase Selectivity and Stability
Post-learning, individual neurons became highly specialized, firing only during specific phases of the movement cycle. This behavior closely mirrors the activity of real spinal interneurons. Quantitatively, the system showed high stability; while pre-learning movements decayed due to friction, post-learning trials consistently resulted in non-decaying rhythms.
Discussion
This study demonstrates that rhythmic movement can emerge from the interaction between a simple neural network and physical mechanics. By using a universal computational principle like the BCM rule, the model achieves a level of adaptability that fixed CPG models lack.
Limitations and Future Directions
The model currently relies on a manually tuned "force factor" to scale neural output to torque, a parameter that would likely be learned in biological systems. Additionally, while pendulums provide a foundational proof of concept, future work should explore more complex musculoskeletal dynamics. The authors suggest that this model does not replace CPG theory but rather complements it, showing how learned connectivity can interact with genetic structures to produce robust, adaptive locomotion.
Summary updated