The Co-Contraction Channel: Using Antagonist Muscle Balance as a Direct Conditioning Signal for Generative Movement

What if the AI could read the difference between a held gesture and a released one — directly from the muscles, in real time?

The Provocation

Two dancers perform the same arm gesture. On video, on motion capture, on any skeletal recording, the two gestures are identical: the same path, the same speed, the same endpoint. Yet anyone watching in person perceives them as completely different. One arm is held — controlled, contained, ready to stop or reverse at any instant. The other is released — flowing, committed, carried by its own momentum.

This is Laban's distinction between bound and free Flow, and it is one of the most fundamental qualitative dimensions in all of movement practice. It is also, until now, essentially invisible to AI motion systems, because it does not show up in the skeletal trajectory.

But it is not invisible to electromyography. The physical difference between a bound and a free movement is, to a large degree, a difference in co-contraction: the simultaneous activation of antagonist muscle pairs. A bound movement co-contracts (the bicep and tricep fire together, stiffening and controlling the joint); a free movement does not (the antagonist relaxes, letting the movement flow).

This brief proposes building a generative movement system conditioned directly on a real-time co-contraction signal — making the bound/free dimension a controllable input rather than an invisible quality.

The Physical Basis

Co-contraction is one of the best-understood phenomena in motor control. When antagonist muscle pairs (e.g., bicep/tricep, quadriceps/hamstring) activate simultaneously, the joint they span becomes stiffer and more controlled; the movement becomes slower to perturb and easier to halt or reverse. When the antagonist relaxes during agonist activation, the joint is compliant and the movement flows with less moment-to-moment control.

The co-contraction ratio — the ratio of antagonist to agonist activation — is directly measurable from surface EMG on the two muscles of an antagonist pair. It is a single, continuous, low-dimensional value that captures much of what distinguishes bound from free movement:

  • High co-contraction ratio → bound, held, controlled, braced (Laban: bound Flow)
  • Low co-contraction ratio → free, released, flowing, committed (Laban: free Flow)

This is a rare case of a clean mapping between a measurable physiological signal and a fundamental somatic quality dimension. Most quality dimensions are multiply determined and hard to isolate. Co-contraction → Flow is unusually direct.

The Proposed System

Sensing: A minimal EMG configuration — two channels per monitored joint, placed on an antagonist muscle pair (e.g., bicep/tricep for the elbow). For a basic upper-body system, 4 antagonist pairs (both elbows, both shoulders) = 8 EMG channels. Sampling at 1,000 Hz; co-contraction ratio computed in rolling 50ms windows.

The co-contraction channel: From each antagonist pair, compute a single continuous co-contraction ratio. The set of ratios across monitored joints forms a compact "Flow vector" — a real-time, per-joint readout of how bound or free each part of the body currently is.

Conditioning: The Flow vector is injected as a conditioning signal into a generative movement model — using the parameter-efficient injection approach now standard in the field (cf. KV-Control, June 8 news digest; REACT FiLM conditioning, June 2 news digest). The base generative model is frozen; the Flow vector modulates its output through a compact injection interface.

Generation: The model generates visual or avatar movement whose Flow quality tracks the practitioner's measured co-contraction. When the practitioner's movement becomes bound, the generated movement becomes bound; when they release into free flow, the generated movement releases with them — at the level of quality, not just trajectory.

Why this is newly feasible: Two developments from this month make this practical. First, muscle-level motion models (MuscleMimic, June 15 news digest) validated against EMG demonstrate that the muscle-activation-to-movement-quality relationship can be modelled computationally. Second, the maturation of parameter-efficient conditioning (KV-Control, REACT) means the Flow vector can be added to a powerful frozen generative model without retraining.

The Experiment

Setup: A single practitioner wears the 8-channel antagonist-pair EMG configuration. Two conditions, compared across a structured improvisation:

  1. Trajectory-only conditioning (control): generated movement conditioned on the practitioner's joint positions only.
  2. Trajectory + Flow conditioning (experimental): generated movement conditioned on joint positions and the real-time co-contraction Flow vector.

Primary measure: Trained observers (Laban Movement Analysts) rate, blind to condition, how well the generated movement's Flow quality matches the practitioner's. Hypothesis: the experimental condition produces matches that LMA-trained observers rate as substantially more faithful in the bound/free dimension.

Secondary measure: The practitioner's own report of whether the generated movement "felt like it understood" the held/released quality of their movement — the felt sense of being met at the level of quality rather than position.

Falsifiable prediction: If co-contraction conditioning works as hypothesised, the experimental condition should show its largest advantage precisely on movements where trajectory is held constant but Flow varies — the same arm gesture performed bound vs. free. On these matched-trajectory pairs, the control condition (trajectory-only) cannot distinguish them at all; the experimental condition should track the difference cleanly.

Why This Has Not Been Done

The technical pieces have existed for years — EMG co-contraction measurement is decades old, and generative movement models are now mature. The reason the combination has not been built is the same conceptual barrier identified in previous briefs: the field's default assumption that movement is its trajectory, and that quality is either decorative or derivative.

The co-contraction channel inverts this. It treats the bound/free quality dimension as a primary, controllable input — as fundamental to the movement as its path through space. This is a somatic premise (quality is primary) implemented as a technical architecture (quality is a conditioning channel).

What Success Would Establish

If the co-contraction channel works, it would be the first demonstration that a fundamental somatic quality dimension can be sensed, transmitted, and generatively rendered in real time — that the felt difference between a held and a released gesture, invisible to every skeletal system, can be made legible to an AI through the muscular signal that produces it.

It would be a single clean proof of the larger SSIN thesis: that sensing at the muscular layer gives an AI access to the quality dimension where somatic practice lives, and that a co-creative dialogue about quality — not just position — becomes possible.

One quality dimension (Flow, via co-contraction) is a modest, achievable first target. But it is the right first target: clean mapping, minimal sensing, falsifiable prediction, and a direct line to the central claim.

Supporting Literature

Bernstein, N. A. (1967). The co-ordination and regulation of movements. Pergamon Press.

Hogan, N. (1984). Adaptive control of mechanical impedance by coactivation of antagonist muscles. IEEE Transactions on Automatic Control, 29(8), 681–690. https://doi.org/10.1109/TAC.1984.1103644

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

Sun, T., et al. (2026). KV-Control: Parameter-efficient K/V injection for trajectory-controlled text-to-motion. arXiv:2606.05624. https://arxiv.org/abs/2606.05624