Bridging the Isometric to Dynamic Motor Control Gap

30 concepts28 papers1 proposal

To understand how the brain controls movement, neuroscientists and engineers have historically relied on a convenient simplification: the isometric contraction. In an isometric task, a participant exerts force against a stationary sensor. Because the limb remains perfectly still, the geometric relationship between muscles and recording electrodes stays constant, preventing motion artifacts and signal distortion. This experimental stability has allowed researchers to decode individual motor neuron spikes and map out the fundamental principles of motor control.

However, this convenience comes at a steep cost. Real-world movement is inherently dynamic. When we walk, reach, or cycle, muscles rapidly change length, joints rotate, and the mechanical properties of the musculoskeletal system shift continuously. Models and decoders trained under static conditions routinely break down when transitioned to dynamic tasks. This discrepancy—the isometric-to-dynamic gap—is one of the most significant hurdles in modern motor neuroscience and neuroengineering.

Do Static Synergies Hold in Motion?

A cornerstone of motor control theory is that the central nervous system (CNS) simplifies the control of our complex bodies by grouping motor units into low-dimensional "neural synergies" or "common drives." For instance, Tanzarella et al. demonstrated in Synergistic Organization of Neural Inputs from Spinal Motor Neurons to Extrinsic and Intrinsic Hand Muscles that thousands of individual motor neurons in the hand are organized into just a few highly consistent neural synergies. Similarly, Laine et al. found that synergistic thigh muscles share over 80% of their common synaptic input during static tasks (Motor Neuron Pools of Synergistic Thigh Muscles Share Most of Their Synaptic Input).

But do these low-dimensional organizational rules persist when the body actually moves? As Levine et al. note in Two motor neuron synergies, invariant across ankle joint angles, activate the triceps surae during plantarflexion, while these synergies may remain stable across different static joint angles, it remains an open question whether they survive continuous changes in muscle length and joint velocity.

Furthermore, recent evidence suggests that the apparent rigidity or flexibility of motor unit recruitment might be task-dependent. While some static tasks reveal highly rigid recruitment patterns constrained by the size principle (Del Vecchio et al. in Interfacing Motor Units in Nonhuman Primates Identifies a Principal Neural Component for Force Control Constrained by the Size Principle), others reveal high-dimensional, flexible control strategies (Marshall et al. in Flexible neural control of motor units). Because these high-dimensional observations are restricted to stationary limbs to ensure signal stability, we do not yet know if the brain deploys entirely different organizational rules during fluid, kinematic displacement.

Shifting Physics and Sensor Geometry

Why exactly do models trained on static data fail during dynamic movement? The breakdown happens at both the mechanical level and the sensor interface.

From a mechanical perspective, translating neural firing into physical force is relatively straightforward when joint angles are fixed. However, during dynamic movement, the force a muscle fiber can generate depends non-linearly on its instantaneous length and velocity (the force-length and force-velocity relationships). As Sartori demonstrates in In Vivo Neuromechanics: Decoding Causal Motor Neuron Behavior with Resulting Musculoskeletal Function, while open-loop prediction of torque from motor neuron spikes is highly feasible under isometric conditions, accounting for rapidly shifting muscle fiber dynamics during real-world movement introduces immense computational complexity.

From a sensor perspective, non-invasive techniques like high-density electromyography (HD-EMG) rely on surface electrodes remaining perfectly aligned with the underlying muscle fibers. During dynamic movement, muscles change shape, bulge, and slide beneath the skin. This physical displacement alters the source of the electrical signals relative to the sensors. The authors of Non-invasive analysis of motor neurons controlling the intrinsic and extrinsic muscles of the hand point out that this shifting geometry severely complicates the stable decomposition of surface signals into reliable neural commands. Consequently, neural network models optimized for static contractions exhibit significantly higher error rates when predicting torque during dynamic motion (Chandrapal et al. in Investigating improvements to neural network based EMG to joint torque estimation).

Task-Dependent Neural Tuning

The gap is not merely a technical artifact of sensors or muscle physics; the central nervous system itself appears to alter its coordination strategies depending on whether a task is static or dynamic.

Consider how muscles synchronize. Mohr et al. investigated this in Task-Dependent Intermuscular Motor Unit Synchronization between Medial and Lateral Vastii Muscles during Dynamic and Isometric Squats, finding that the brain dynamically tunes the synchronization of quadriceps muscles based on the task. While the muscles synchronize strongly to generate torque during dynamic squats, they decouple to act independently during static balance tasks. This task-dependent decoupling means that static observations fundamentally fail to capture the neural demands of movement.

Similarly, the frequency-specific communication channels between the brain and muscles—such as corticomuscular coherence and beta-band bursts—are highly sensitive to state transitions. Studies on these oscillatory signals are traditionally confined to isometric contractions because movement introduces massive artifacts (Liu et al. in Corticomuscular Coherence and Its Applications: A Review). Yet, as explored by Borzelli et al. in Frequency-specific intermuscular coherence of synergistic muscles during an isometric force generation task and Graef et al. in Propagation of Beta bursts from the motor cortex to the motor units of multiple upper-limb muscles, these transient neural oscillations may modulate entirely differently during fluid, real-world transitions than during sustained static force production.

Strategies for Generalization

To overcome the isometric constraint, researchers are attacking the problem from two sides: robust deep learning and physiologically realistic modeling.

On the machine learning front, the goal is to build architectures that can inherently generalize across changing physical contexts. For example, Wen et al. developed a deep convolutional neural network in Toward a generalizable deep CNN for neural drive estimation across muscles and participants designed to maintain stable neural drive estimation despite the muscle shape changes and electrode shifts characteristic of dynamic tasks.

On the computational side, frameworks like the one proposed by Caillet et al. in Estimation of the firing behaviour of a complete motoneuron pool by combining electromyography signal decomposition and realistic motoneuron modelling combine experimental data with realistic mathematical models of entire motor neuron pools. By baking the laws of muscle physics directly into the decoder, these hybrid frameworks aim to preserve predictive accuracy even when muscle properties change rapidly during locomotion.

Ultimately, moving from the clean, controlled world of isometric force to the messy reality of dynamic movement is not just a matter of scaling up existing models. It requires accounting for a fundamental shift in how the nervous system coordinates, how sensors capture electrical activity, and how the physical body translates those commands into action.

Key Foundational Work

Investigating improvements to neural network based EMG to joint torque estimation (2011) — contributed 1 concept elements to this direction.
Electromyographic Study on the Effects of Head Position to the Activities of Head and Neck Muscles (1990) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 28 papers
  2. Version 1 · 2026-10-04 · 44 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_isometric_dynamic_generalizability_c79e01c6,
  title = {isometric-dynamic generalizability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/isometric-dynamic-generalizability/rdr_c79e01c6069c44cbafcff0be1b84eeac},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_isometric_dynamic_generalizability_c79e01c6_v2,
  title = {isometric-dynamic generalizability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/isometric-dynamic-generalizability/rdr_c79e01c6069c44cbafcff0be1b84eeac/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake