The Pre-Movement Window: An EMG Pre-Activation Interface for Anticipatory Somatic-AI Conditioning
What if the AI responded to your intention, not your action?
The Provocation
Every movement you make is preceded by a signal you cannot consciously observe: an electromyographic (EMG) burst in the muscles involved, firing 50 to 200 milliseconds before visible movement begins. This is the pre-activation window — the nervous system's instruction to the body that precedes its execution.
Movement science has known about this window for decades. It appears reliably in skilled movers, it carries information about the quality and direction of the upcoming movement, and it is accessible via surface EMG sensors placed on the skin. What movement science has not yet done is use this signal as the primary conditioning input for a generative AI system.
This brief proposes that experiment.
What the Pre-Activation Window Reveals
Surface EMG detects electrical potential generated by muscle fibre recruitment. In the pre-activation window, the pattern of which muscles activate, in what order, and at what intensity provides the following information before any visible movement occurs:
- Direction — the pre-activation pattern for a forward reach differs from a lateral reach; the body is already committing before the arm moves.
- Effort quality — a light touch and a firm push involve different pre-activation amplitudes and temporal profiles, detectable before contact.
- Intention strength — pre-activation amplitude correlates with the force of the intended movement; tentative movements show lower pre-activation than decisive ones.
- Preparatory tension — co-contraction patterns (multiple muscle groups firing together before movement) indicate a held, braced, or cautious intention; sequential patterns indicate a fluid, released one.
Somatic practitioners develop sensitivity to exactly these qualities through training — not via EMG readout, but through cultivated proprioceptive attention. The pre-activation window is the electrophysiological correlate of what the somatic practitioner learns to sense.
The Proposed Experiment
Condition: A single practitioner wears a 16-channel surface EMG array (covering the major muscle groups of torso, arms, and upper legs). EMG is sampled at 2,000 Hz — 16× faster than standard motion capture. A latency-optimised signal processing pipeline extracts the pre-activation envelope in rolling 25ms windows.
Conditioning pathway: The pre-activation envelope is encoded into a low-dimensional vector (8–16 features: per-channel RMS amplitude, onset timing, co-contraction ratio). This vector is passed as a conditioning signal to a generative model trained on the practitioner's movement-visual pairs.
Generation: The generative model produces visual output conditioned on the current pre-activation vector. Crucially, this output begins updating before any visible movement has occurred — it is conditioned on the body's anticipatory signal, not on the resulting action.
Evaluation: Two conditions compared across a 20-minute improvisation session:
- Post-movement conditioning (control): generative output conditioned on joint positions extracted from video at standard frame rate.
- Pre-movement conditioning (experimental): generative output conditioned on EMG pre-activation envelope.
Primary outcome measure: temporal alignment between practitioner's felt movement intention and onset of generative output change. Secondary measure: practitioner's qualitative assessment of "dialogue quality" — whether the visual output felt responsive or reactive.
Why This Has Not Been Done
The technical barriers are modest; the conceptual barrier is significant.
Most somatic-AI interaction systems default to vision as the sensing modality because cameras are accessible, interpretable, and produce data formats that connect directly to existing motion generation pipelines. EMG requires skin contact, electrode placement expertise, signal processing knowledge, and the willingness to work with a noisier, more embodied data source.
The deeper barrier is conceptual: the assumption that AI systems should respond to what the body does, not what it intends to do. This assumption is built into every system evaluated by comparing generated motion to recorded motion — the ground truth is always the completed action, not the anticipatory signal that preceded it.
Somatic practice inverts this priority. The intention is the primary material; the visible movement is its expression in the world. A co-creative AI system designed for somatic practice should, by design, respond to intention rather than action.
What Success Would Look Like
A practitioner engaged with this system would experience the generative visual output as predictive of their movement rather than reactive to it. Rather than watching the visual respond after they have moved, they would feel the visual begin to move with them — or even slightly before them — as their nervous system commences its pre-activation sequence.
This is not magic. It is a consequence of placing the conditioning signal upstream of visible movement, in the pre-activation window. The generative system would be doing exactly what it is trained to do: condition on available input and generate an appropriate output. The input has simply been moved 50–200ms earlier in the movement cycle.
The qualitative consequence — the feeling of being anticipated rather than tracked — is precisely what distinguishes a creative partnership from a mirror.
Supporting Literature
Cram, J. R., Kasman, G. S., & Holtz, J. (1998). Introduction to surface electromyography. Aspen.
Godinho, M., Meira, C., & Mendes, L. (2008). Anticipatory postural adjustments in motor control: A review. Motor Control, 12(3), 199–221.
Narayanan, S., Jiang, Z., Narasimhan, S., & Chandraker, M. (2026). PhyCo: Learning controllable physical priors for generative motion. arXiv:2604.28169. https://arxiv.org/abs/2604.28169
Zhang, X., et al. (2026). PersonaGesture: Single-reference co-speech gesture personalization for unseen speakers. arXiv:2605.06064. https://arxiv.org/abs/2605.06064