Information Based Centralization of Locomotion in Animals and Robots

TL;DR

This paper introduces new information-theoretic measures to quantify control centralization in biological and robotic locomotion, validating them through simulations, animal experiments, and robotic models to understand coupling mechanisms.

Contribution

It develops model-free measures of control centralization and co-information, applying them to biological and robotic systems to reveal insights into coupling and control strategies.

Findings

Cockroach leg control shows unexpected centralization in middle leg muscles.

Centralization correlates with stride frequency but remains consistent across groups.

Mechanical coupling influences control centralization independently of neural control.

Abstract

Movement in biology is often achieved with distributed control of coupled subcomponents, e.g. muscles and limbs. Coupling could range from weak and local, i.e. decentralized, to strong and global, i.e. centralized. We developed a model-free measure of centralization that compares information shared between control signals and both global and local states. A second measure, called co-information, quantifies the net redundant information the control signal shares with both states. We first validate our measures through simulations of coupled oscillators and show that it successfully reconstructs the shift from low to high coupling strengths. We then measure centralization in freely running cockroaches. Surprisingly, extensor muscle activity in the middle leg is more informative of movements of all legs combined than the movements of that particular leg. Cockroach centralization successfully recapitulates a specific model of a strongly coupled oscillator network previously used to model cockroach leg kinematics. When segregated by stride frequency, slower cockroach strides exhibit more shared information per stride between control and output states than faster strides, indicative of an information bandwidth limitation. However, centralization remains consistent between the two groups. We then used a robotic model to show that centralization can be affected by mechanical coupling independent of neural coupling. The mechanically coupled bounding gait is decentralized and becomes more decentralized as mechanical coupling decreases while internal parameters of control remain constant. The results of these systems span a design space of centralization and co-information that can be used to test biological hypotheses and advise the design of robotic control.