What if the AI didn't just move with you — but moved as if it could feel you about to move?
The Provocation
The real-time barrier has fallen. This month, generation at 33ms latency (ARDY) brought machine movement inside the window where the human nervous system experiences a response as simultaneous rather than delayed. Reactive-interaction research is producing systems that generate one mover's response conditioned on another's. The technical substrate for an AI movement partner — fast enough, reactive enough — now exists.
But a movement partner that reacts to your visible movement, however fast, is not yet a movement partner in the sense that Contact Improvisation or any deep partnering practice means. Real partners do not react to what you have done. They anticipate what you are about to do, so that the two of you arrive at the shared moment together rather than one following the other. The felt signature of genuine partnering is not fast response — it is mutual anticipation.
This brief proposes building the first anticipatory movement partner: a reactive generation system conditioned not on the practitioner's visible movement but on their pre-movement intention, sensed through EMG pre-activation — so that the generated partner responds to the movement as it forms, not after it appears.
Why Reaction Is Not Enough
Consider the difference, in Contact Improvisation, between a beginner and an experienced partner receiving your weight.
The beginner reacts. You lean; they feel the lean; they respond. Even when they respond quickly, there is a felt gap — a sense of you initiating and them catching up. The contact works, but it has the quality of action-and-response.
The experienced partner anticipates. As your weight begins to shift — before it is a visible movement, while it is still forming in the preparatory organisation of your body — they are already meeting it. The result is a felt togetherness: you arrive at the shared weight together, neither leading nor following. This is the quality partnering practice cultivates, and it is categorically different from fast reaction.
The difference is temporal placement relative to the movement's formation. Reaction responds to the movement once it exists. Anticipation responds to the movement as it forms — which requires access to the forming, the pre-movement organisation, not just the formed visible result. As this month's synthesis argued, speed gets a response inside the window of felt simultaneity, but only anticipation produces genuine co-movement. A 33ms reaction to visible movement is still a reaction. An anticipatory response to forming movement is a partnership.
The Signal That Makes Anticipation Possible
The pre-movement organisation of a movement is not visible — but it is measurable. Surface EMG detects muscle activation 50–200ms before visible movement begins (the pre-activation window, detailed in the May innovation brief). This anticipatory signal carries information about the direction, effort, and quality of the movement that is about to occur, before it occurs.
This is precisely the signal a human partner is implicitly reading — not via EMG, but through trained proprioceptive attunement to a partner's forming movement. The experienced partner feels your weight-shift beginning in the subtle preparatory changes in your body before it becomes a visible lean. EMG pre-activation is the measurable correlate of exactly this forming-movement information.
An AI conditioned on the practitioner's EMG pre-activation, rather than their visible movement, would have access to the forming movement — the same information the anticipating partner reads. It could respond to the movement as it forms, arriving at the shared moment together rather than after.
The Proposed System
Sensing: The practitioner wears an EMG array (building on the co-contraction configuration from the June brief — 8 antagonist-pair channels, extensible). The pre-activation envelope is extracted in rolling windows, yielding a real-time signal of forming movement 50–200ms ahead of visible motion.
Prediction: The pre-activation signal is decoded into an anticipated movement vector — a short-horizon prediction of the direction, effort, and quality of the movement about to occur. This is not the movement itself but its forming intention.
Reactive generation: A real-time reactive generation backbone (ARDY-class, sub-window latency; contact-aware where partnering involves contact, per Contact Matrix) generates a partner's movement conditioned on the anticipated movement vector — responding to the forming movement rather than the visible one.
The temporal result: Because the conditioning signal precedes visible movement by 50–200ms, and the generation latency is ~33ms, the generated partner's response can be produced before or exactly as the practitioner's visible movement occurs — achieving genuine anticipatory co-movement rather than fast reaction.
Evaluation: A Contact Improvisation practitioner engages the system in two conditions:
- Reactive (control): partner generated from the practitioner's visible movement (position/velocity), sub-window latency.
- Anticipatory (experimental): partner generated from EMG pre-activation (forming movement).
Primary measure: The practitioner's felt assessment of partnership quality — specifically, whether the generated partner feels anticipatory (arriving together) or reactive (catching up). This is the categorical distinction partnering practitioners are trained to feel.
Falsifiable prediction: The experimental (anticipatory) condition should produce the felt quality of "arriving together" that the reactive condition — however low its latency — cannot, because reaction to visible movement is structurally after-the-fact regardless of speed. If low-latency reaction alone produced the anticipatory felt quality, the pre-movement conditioning would show no advantage; the prediction is that it will, specifically on the arriving-together dimension.
Why This Has Not Been Done
Every component now exists: real-time reactive generation (this month), contact-aware interaction modelling (this month), EMG pre-activation sensing (decades old), fast decoding (mature). What has not been done is the combination — and the reason is conceptual, not technical.
The field's reactive-interaction research conditions on observed partner movement, because that is the available signal in the standard setup (two agents observing each other visually). The move to conditioning on pre-movement intention requires the somatic premise that the forming movement, not the formed one, is the ground of partnership — and the sensing modality (EMG) that accesses it. This is the somatic contribution to the reactive-generation frontier: not faster reaction, but response to the forming rather than the formed.
What Success Would Establish
If the anticipatory duet works, it would demonstrate the specific thing that real-time reaction alone cannot: that an AI can produce the felt quality of being anticipated — of a partner arriving with you rather than after you — by responding to your movement as it forms.
This is the qualitative heart of partnering practice, and it is exactly the dimension that fast-but-reactive systems, no matter how low their latency, structurally cannot reach. It would establish that somatic sensing (pre-movement EMG) contributes something categorical, not incremental, to interactive movement AI: the difference between a fast mirror and a genuine partner.
It would also be an interaction-stage existence proof: an anticipatory movement partner that builds on earlier sensing and conditioning work.
Continue through the archive
For connected context, read The Window of Now: Interpersonal Temporal Binding and the Latency Threshold of Responsive Generation and Thinking in Movement: Maxine Sheets-Johnstone and the Intent Layer as a Site of Cognition, Not Execution.
References
Repp, B. H., & Su, Y.-H. (2013). Sensorimotor synchronization: A review of recent research (2006–2012). Psychonomic Bulletin & Review, 20(3), 403–452. https://doi.org/10.3758/s13423-012-0371-2
Godinho, M., Meira, C., & Mendes, L. (2008). Anticipatory postural adjustments in motor control: A review. Motor Control, 12(3), 199–221.
Zhao, K., Petrovich, M., Zhang, H., Wang, T., Tang, S., & Rempe, D. (2026). ARDY: Autoregressive diffusion with hybrid representation for interactive human motion generation. arXiv:2607.08741. https://arxiv.org/abs/2607.08741
Contact Matrix: Enhancing dance motion synthesis with precise interaction modeling. (2026). arXiv:2605.04662. https://arxiv.org/abs/2605.04662