Muscles Reduce Neuronal Information Load: Quantification of Control Effort in Biological vs. Robotic Pointing and Walking

Daniel F. B. Haeufle, Isabell Wochner, David Holzmüller, Danny Driess, Michael Günther, Syn Schmitt

TL;DR

This paper introduces a novel algorithm to quantify "control effort" – the minimal information needed to perform a movement – and applies it to compare biological (muscle-driven) and robotic (torque-driven) systems for pointing and walking. The findings consistently show that muscle-driven systems require significantly less control effort, suggesting that the inherent properties of muscles contribute to efficient biological control.

What Problem Does This Paper Address?

The nervous system faces the challenge of controlling complex biological systems with inherent nonlinearities and delays. A prevailing hypothesis is that the "morphology" of these systems – their physical structure and material properties – plays a crucial role in simplifying control, a concept known as "morphological computation." This means that the physical system itself performs some of the computational work, reducing the burden on the central nervous system. However, quantifying this benefit, or "control effort," has been challenging, especially for complex movements. Previous work quantified control effort for simple movements, but this paper extends the methodology to more realistic human arm pointing and walking, tackling the problem of how to efficiently measure control effort in systems with many control signals.

Methods

The core of this paper is the development of a new algorithm to quantify control effort. Control effort is defined as the minimum information entropy required to achieve a desired movement, subject to a performance constraint. The algorithm is a derivative-free direct search method based on the pattern search concept, specifically designed to handle the computationally expensive nature of evaluating the performance constraint. It involves three phases: a rapid parallel reduction of resolution, a more thorough pattern search, and a check of the local neighborhood.

For evaluating control effort, the authors used numerical simulations of two complex human movements:

  • Pointing: A 2D arm model with either six Hill-type muscle-tendon units or two ideal torque actuators. Controllers were learned using deep reinforcement learning.
  • Walking: A seven-segment multi-body model with 14 muscles or six ideal torque actuators. The muscle-driven model used a reflex-based neural control scheme, while the torque-driven model used PD controllers.

The control effort was quantified by discretizing control signals (muscle stimulations or torque commands) and sensor signals, and then optimizing the resolution parameters to minimize information content while maintaining performance.

Datasets

The paper does not introduce new datasets. Instead, it utilizes existing musculo-skeletal models for arm movements and walking, which have been previously used in other research. The models themselves, along with their parameters, are described in the supplementary material.

Key Insights and Ideas

  • Morphology as a Computational Offload: The study provides strong quantitative evidence that the inherent properties of biological muscles significantly reduce the "control effort" compared to artificial torque actuators. This supports the idea that biological evolution has leveraged physical properties to simplify neural control.
  • System Property, Not Controller Property: The observed reduction in control effort for muscle-driven systems was consistent across different controllers (reinforcement learning and reflex-based) and movements (pointing and walking). This suggests that the benefit is a fundamental property of the neuromuscular system itself, rather than a specific controller design.
  • Novel Algorithm for Complex Systems: The developed pattern search-based algorithm is crucial for enabling the quantification of control effort in systems with a high number of control signals, overcoming the limitations of brute-force methods used previously.
  • Sensory vs. Motor Discretization: For pointing, discretizing muscle stimulations (motor output) resulted in lower control effort than discretizing sensor inputs, suggesting that the output stage of the neuromuscular system is particularly efficient.
  • Delay's Role: In biological systems, delays (like electromechanical delay) can actually reduce control effort, contrary to what might be expected in engineering. This highlights how biological systems exploit their inherent dynamics.

Key Results

  • Pointing: The muscle-driven STIM scenario (discretizing muscle stimulations) had the lowest control effort (67.5 bit/s). The SENS scenario (discretizing sensor inputs) required about three times more effort (222.5 bit/s), and the torque-driven model required almost four times more (310.3 bit/s).
  • Walking: Both STIM and SENS scenarios for walking required about half the control effort of the best torque-driven model.
  • Algorithm Efficiency: The proposed algorithm converged much faster and found better solutions compared to standard optimization techniques like dual annealing and differential evolution for the pointing task.
  • Error Estimation: The estimated error in control effort (ΔIopt) was small, indicating the reliability of the optimization results.

Limitations and Future Work

  • Idealized Models: The study uses simplified models of biological systems and idealized torque actuators. Real-world robots and biological systems are more complex.
  • Specific Movements: The analysis is limited to pointing and walking. Other movements might yield different results.
  • "Minimal" Information vs. Actual Processing: The paper acknowledges that biological systems likely do not operate at the absolute minimum information limit, as this would be on the verge of instability. The "control effort" quantifies the potential reduction due to morphology.
  • Future Work: The authors suggest applying the algorithm to other robotic systems and exploring how different morphological characteristics influence control effort. They also propose investigating other optimal control approaches like "Minimum Attention Control" (MAC) for measuring simplicity and comparing it to control effort. The influence of delays on control effort in biological systems warrants further investigation.

Summary updated

Abstract

It is hypothesized that the nonlinear muscle characteristic of biomechanical systems simplify control in the sense that the information the nervous system has to process is reduced through off-loading computation to the morphological structure. It has been proposed to quantify the required information with an information-entropy based approach, which evaluates the minimally required information to control a desired movement, i.e., control effort. The key idea is to compare the same movement but generated by different actuators, e.g., muscles and torque actuators, and determine which of the two morphologies requires less information to generate the same movement. In this work, for the first time, we apply this measure to numerical simulations of more complex human movements: point-to-point arm movements and walking. These models consider up to 24 control signals rendering the brute force approach of the previous implementation to search for the minimally required information futile. We therefore propose a novel algorithm based on the pattern search approach specifically designed to solve this constraint optimization problem. We apply this algorithm to numerical models, which include Hill-type muscle-tendon actuation as well as ideal torque sources acting directly on the joints. The controller for the point-to-point movements was obtained by deep reinforcement learning for muscle and torque actuators. Walking was controlled by proprioceptive neural feedback in the muscular s...
BibTeX
@article{haeufle2020muscles,
  author = {Daniel F. B. Haeufle and Isabell Wochner and David Holzmüller and Danny Driess and Michael Günther and Syn Schmitt},
  title = {{Muscles Reduce Neuronal Information Load: Quantification of Control Effort in Biological vs. Robotic Pointing and Walking}},
  journal = {Frontiers Robotics AI},
  year = {2020},
  doi = {10.3389/frobt.2020.00077},
}
Figure