Morphology based Locomotion Controller for Modular Robots

Overview

A distributed neural control approach for generating locomotion in linear modular robots. Each module runs the same small neural network, using only local connectivity and neighboring actuator states to produce coordinated oscillatory motion. The controller is evolved with a Genetic Algorithm and tested on simulated Y1 modular robots, demonstrating stable gait generation, frequency-adaptive actuation, and recovery from external perturbations.

Key Concepts

• Proposes a simple per-module MLP (5–5–1) whose oscillatory closed-loop output drives each module actuator, yielding a fully distributed controller.
• Makes the controller homogeneous (same topology + weights across modules) while achieving coordination via module-specific inputs (connectivity + neighbor actuator states).
• Uses a standard GA (roulette selection + recombination + mutation) to optimize synaptic weights for stable locomotion gaits.
• Introduces and compares two actuation schemes: fixed-interval neural actuation vs frequency-adaptive actuation triggered when joints reach target angles (improving speed).
• Demonstrates fault tolerance / limit-cycle recovery by evolving controllers evaluated from random initial conditions and validating recovery under injected perturbations in simulation (OpenRAVE/Y1).