Below the Skeleton: Muscle Redundancy and the Somatic Concept of Movement Quality as Convergent Discoveries
Confidence Label: Well-Supported
Connection type: Structural correspondence between a biomechanical property (muscle redundancy) and a somatic-pedagogical concept (movement quality) — independently established in both domains, now meeting at the muscle-level AI frontier
The Synthesis
Biomechanics and somatic practice have independently arrived at the same insight from opposite directions: the same visible movement can be produced in many different ways, and the difference between those ways is where the meaning lives.
In biomechanics, this is the problem of muscle redundancy (also called motor redundancy or the degrees-of-freedom problem, originally formulated by Nikolai Bernstein): the human body has far more muscles than strictly necessary to achieve any given joint trajectory, so any movement can be produced by an infinite family of muscle activation patterns. The skeleton's path through space radically underdetermines the muscular activity that produced it.
In somatic practice, the corresponding insight is the concept of movement quality: two movers can trace identical paths through space — the same arm raise, the same step, the same turn — while the quality of those movements differs completely. One is effortful, one is easy; one is held, one is released; one expresses anxiety, one expresses ease. The visible trajectory is the same; the quality is different.
These are the same observation. Muscle redundancy is the biomechanical fact that underlies the somatic phenomenon of movement quality. The newly tractable muscle-level AI models arriving this month (MuscleMimic and related work) are, for the first time, building computational systems at the level where this convergence becomes operational.
Muscle Redundancy: The Biomechanical Fact
The human body contains roughly 600 muscles. The skeletal degrees of freedom they control number in the low hundreds. This mismatch is not an inefficiency — it is a fundamental design feature, and it has a profound consequence: the mapping from muscle activation to movement is many-to-one. Many different patterns of muscle activation produce the same joint trajectory.
Nikolai Bernstein, the Soviet physiologist who first formalised this in the mid-twentieth century, recognised that this redundancy poses a control problem (how does the nervous system choose among the infinite options?) but also affords a richness: the same goal can be achieved with different muscular "styles," different distributions of effort, different relationships between agonist and antagonist muscles.
Two especially important dimensions of this redundancy for our purposes:
Co-contraction. The body can move a joint by activating the agonist muscle while relaxing the antagonist (an efficient, "free" movement), or by activating both simultaneously (a stiff, "bound," controlled movement). The joint trajectory can be identical; the co-contraction level — and therefore the quality — differs completely. This maps directly onto Laban's Flow effort (free vs. bound).
Effort distribution. The same lift can be performed with effort concentrated in large proximal muscles (a grounded, powerful organisation) or distributed toward smaller distal muscles (a lighter, more delicate organisation). Again: same visible result, different muscular production, different felt and perceived quality.
These are not subtle laboratory findings. They are measurable with EMG, and they correspond precisely to the qualitative distinctions that somatic practitioners are trained to perceive and produce.
Movement Quality: The Somatic Concept
Somatic practice is, in large part, the cultivation of sensitivity to and control over movement quality — the dimension of movement that is independent of its spatial trajectory.
A somatic educator watching a student perform a simple arm raise perceives far more than the path the arm traces. They perceive whether the movement is initiated from the core or the periphery, whether it carries unnecessary holding in the shoulder, whether the breath supports or interrupts it, whether the quality is anxious or settled, effortful or easy. These perceptions are not vague impressions — they are reliable, teachable discriminations that practitioners can name, reproduce, and modify.
The entire pedagogy of disciplines like the Alexander Technique, Feldenkrais, Body-Mind Centering, and Laban Movement Analysis rests on the premise that movement quality is real, perceptible, and modifiable independently of movement trajectory. You can teach someone to perform the same visible movement with a different quality — less effort, different initiation, changed flow — and both the mover and a trained observer can verify the change.
What somatic practice has lacked is a precise account of the physical substrate of movement quality. Practitioners know it is real because they can perceive and produce it; but the question of what, physically, distinguishes a high-quality from a low-quality version of the same movement has remained largely phenomenological.
Muscle redundancy provides the answer. The physical substrate of movement quality is the muscular activation pattern — the specific solution, among the infinite redundant possibilities, that the mover's nervous system selects to produce a given trajectory. Movement quality is the muscular organisation of movement. Two arm raises with the same path but different qualities are two different muscular solutions to the same kinematic problem.
Where the Convergence Becomes Operational
Until this year, this convergence was theoretically interesting but not computationally actionable. AI motion systems worked at the skeletal level, where movement quality is, by definition, invisible — the skeleton's trajectory does not encode which of the infinite muscular solutions produced it.
The muscle-level AI models arriving this month change this. MuscleMimic (arXiv:2603.25544) trains policies that generate movement through muscle activation, and validates those activations against real EMG. A model operating at this level is operating at the level where movement quality is physically constituted. For the first time, a computational system represents not just the skeletal trajectory but the muscular solution that produced it — which is to say, the quality.
This has three concrete implications:
1. Movement quality becomes measurable and generable. A muscle-level model can distinguish, and generate, different muscular solutions to the same kinematic goal. The free and bound versions of a movement — identical in trajectory, different in co-contraction — are different outputs of a muscle-level model, where they are indistinguishable to a skeletal model.
2. Somatic quality vocabulary gains a physical grounding. Laban's effort factors, which the May content cluster discussed as an interpretation vocabulary, can be partially grounded in measurable muscular variables: Flow (free/bound) in co-contraction, Weight (strong/light) in activation magnitude and distribution, Time in the temporal profile of activation. This is not a complete reduction — quality is not only muscular — but it is a real physical anchor.
3. The personalisation of quality becomes possible. The CMU exoskeleton work (arXiv:2604.09431) shows muscle-level models can be personalised to an individual's specific patterns, including atypical ones. A practitioner's characteristic movement quality — their signature muscular solutions — becomes representable.
Where the Convergence Is Incomplete
Intellectual honesty requires marking the limits of this synthesis.
Muscle activation is not the whole of quality. Movement quality, as somatic practice understands it, includes dimensions that muscular activation alone does not capture: the relationship of movement to breath, to attention, to intention, to emotional state, to the felt sense of the mover. Muscle redundancy grounds the physical substrate of quality, but quality as a lived phenomenon is richer than its muscular correlate. A complete account would need to integrate the muscular layer with the proprioceptive, attentional, and affective layers.
Measured EMG is not felt quality. As established in the June 1 deep analysis (Fuchs on body memory), there remains an irreducible gap between the third-person measurement of muscular activity and the first-person experience of moving. A muscle-level model validated against EMG has captured the measurable correlate of quality, not the felt quality itself. This is a real and permanent limit of the convergence, not a temporary technical gap.
Redundancy resolution is context-dependent. The nervous system's selection among redundant muscular solutions depends on factors — fatigue, intention, environmental context, history — that a model trained on recorded EMG approximates statistically but does not generate from the same causes. The model learns which solutions occur; it does not undergo the embodied process that selects them.
The Implication for Somatic-AI Co-Creation
This synthesis identifies the most concrete bridge yet between somatic intelligence and AI architecture: the muscular layer is where movement quality is physically constituted, and AI is now able to operate at that layer.
For somatic-AI co-creation, this suggests a specific research direction. An AI system that senses and generates at the muscular level (via EMG sensing and muscle-actuated generation) would be working in the same dimension where the somatic practitioner's expertise lives — the dimension of quality, not just trajectory. The practitioner's lifelong cultivation of muscular discrimination and control, and the AI's muscle-level model, would be commensurable in a way that a practitioner and a skeletal-level model never could be.
This does not close the felt-experience gap. But it locates the dialogue at the right level. A co-creative exchange about movement quality requires both parties to operate in the dimension where quality lives. Muscle-level AI, for the first time, can.
Supporting Evidence
Bernstein, N. A. (1967). The co-ordination and regulation of movements. Pergamon Press.
Latash, M. L. (2012). The bliss (not the problem) of motor abundance (not redundancy). Experimental Brain Research, 217(1), 1–5. https://doi.org/10.1007/s00221-012-3000-4
Li, C., Wang, C., Ziliotto, B., et al. (2026). Towards embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale. arXiv:2603.25544. https://arxiv.org/abs/2603.25544
Choi, I., Park, I., Halilaj, E., & Kang, I. (2026). Musculoskeletal motion imitation for learning personalised exoskeleton control policy in impaired gait. arXiv:2604.09431. https://arxiv.org/abs/2604.09431