Confidence Label: Well-Supported

Connection type: Correspondence between a measured psychophysical constant (the temporal window of felt simultaneity) and an engineering threshold (real-time generation latency) — both governing when a response is experienced as "with me" rather than "after me"

The Synthesis

There is a specific, measurable window of time within which the human nervous system binds events into a felt "now" — and it is the same window that determines whether a generated movement response feels like a partner or a delay.

This month, NVIDIA's ARDY brought real-time motion generation to 33-millisecond latency. That number is not arbitrary from the standpoint of human experience. It sits inside the temporal window within which the nervous system experiences events as simultaneous, within which two movers experience themselves as moving together rather than one-after-the-other, and within which a response is felt as a genuine answer rather than a laggy echo. The engineering threshold and the psychophysical threshold are the same threshold, and understanding why illuminates what real-time responsive generation actually achieves — and what it does not.

The Temporal Window of the "Now"

Human temporal experience is not a knife-edge present but a window — a short span of time within which events are bound together into a single experienced moment. Psychophysical research across the twentieth and twenty-first centuries has characterised this window with some precision.

Several relevant thresholds:

Simultaneity threshold (~20–50ms). Two stimuli separated by less than roughly 20–50ms (depending on modality) are experienced as simultaneous — the nervous system cannot resolve their temporal order and binds them into one event. Above this window, they begin to be experienced as sequential.

The perceptual moment / temporal binding window (~this range and somewhat above). Events falling within this window are integrated into a unified percept. This is the temporal grain of experience — the resolution at which "now" is constructed.

Sensorimotor synchronisation. When people move together — clapping, dancing, walking in step — they synchronise within tens of milliseconds, and crucially they often anticipate rather than react, timing their movement to a predicted beat so that the felt result is togetherness, not follow-the-leader.

The key point for this synthesis: whether an event is experienced as with another event or after it is determined by whether it falls inside or outside this window of tens of milliseconds. Inside the window: simultaneity, togetherness, binding. Outside: sequence, delay, response-after-the-fact.

The Latency Threshold in Responsive Generation

Now the engineering side. A generative system that responds to a human's movement produces its response after some latency — the time from sensing the input to producing the output. Where that latency falls relative to the temporal binding window determines the felt character of the interaction.

  • Latency well above the window (say, 200ms+): The response is clearly experienced as coming after the human's movement. The system is a mirror with a lag — it reflects what you did. However accurate, it cannot feel like a partner, because a partner moves with you, and this system, structurally, moves after you.

  • Latency inside the window (under ~50ms, and better under ~33ms): The response falls within the span the nervous system binds into "now." It is experienced not as a delayed reaction but as part of the same temporal moment — as with the human's movement rather than after it. This is the threshold ARDY crosses.

This is why the sub-50ms figure recurs (this platform's March innovation brief flagged it; ARDY now delivers 33ms). It is not an arbitrary performance target. It is the latency at which a generated response stops being experienced as a reaction and starts being experienced as co-presence. Below the window, the phenomenology of the interaction changes in kind.

Where the Correspondence Holds

The correspondence is genuine and consequential in three respects.

The threshold is real and shared. The temporal binding window is a measured property of the nervous system; the latency budget is a measured property of the system. When the latter falls inside the former, the interaction crosses from felt-sequence to felt-simultaneity. This is not metaphor — it is the same temporal magnitude governing both.

It explains the qualitative shift. Practitioners who have worked with laggy interactive systems and responsive ones report a categorical difference, not a gradual one — a point at which the system stops feeling like a delayed tool and starts feeling like a presence. The binding window explains why the shift is categorical: crossing the window changes the experienced temporal relationship in kind.

It sets a genuine target. For any responsive somatic-AI system, sub-window latency is a real, non-negotiable, and now-achievable requirement. ARDY demonstrates it is reachable. This is a case where the phenomenological requirement and the engineering achievement have met.

Where the Correspondence Breaks Down

As always, the limits are as instructive as the correspondence.

Simultaneity is necessary, not sufficient, for partnership. Falling inside the binding window makes a response feel with rather than after — but co-presence in time is not the whole of partnership. A partner is not merely something that moves simultaneously; it is something that moves simultaneously and appropriately — with the right quality, the right intention, the right responsiveness to what the movement means. A system could hit 33ms and still produce responses that are simultaneous but hollow — present in time, absent in quality. Latency solves the temporal condition of partnership, not its qualitative condition.

Reaction within the window is still reaction. The deepest point. Genuine movement partners do not merely respond fast; they anticipate. In sensorimotor synchronisation, people time their movement to a predicted event, so that the felt result is true togetherness rather than fast-following. A system that senses the human's visible movement and responds within 33ms is still, structurally, reacting — very fast, inside the window, but reacting to movement that has already begun. True co-movement requires anticipating the movement before it is visible — which is exactly what pre-movement sensing (EMG pre-activation) offers and what visible-movement sensing, however fast, cannot. Speed gets you inside the window; anticipation gets you to genuine togetherness.

The window is the practitioner's, not the system's. The binding window is a feature of the human's nervous system experiencing the interaction. The system has no felt "now"; it has a latency budget. The simultaneity is real for the human and merely computational for the machine. This asymmetry is not a flaw to fix but a fact to remember: the felt togetherness lives entirely on the human side, enabled by but not shared with the system.

The Implication for Somatic-AI Co-Creation

This synthesis clarifies what the real-time breakthrough (ARDY) does and does not accomplish for somatic co-creation.

What it accomplishes: it crosses the temporal threshold. A somatic-AI system built on sub-window-latency generation can, for the first time, produce responses that fall within the practitioner's felt "now" — that are experienced as co-present rather than delayed. This is a genuine and necessary achievement. Without it, no amount of quality or intelligence in the response could feel like partnership, because it would always arrive too late to be bound into the same moment.

What it does not accomplish: it does not, by itself, make the response anticipatory or qualitatively appropriate. Crossing the temporal window is the precondition for partnership, not partnership itself. The two further requirements — anticipation (via pre-movement sensing) and qualitative appropriateness (via intent/quality-layer conditioning) — are exactly the contributions the somatic sensing approach offers on top of the real-time substrate.

The picture that emerges is layered. Real-time generation (now achieved) provides the temporal substrate: responses inside the window of now. Anticipatory sensing (EMG pre-activation) provides the anticipation that turns fast reaction into genuine co-movement. Quality-layer conditioning provides the appropriateness that turns simultaneous movement into meaningful partnership. The temporal threshold is the foundation; the somatic contributions are what build partnership on it.

Continue through the archive

For connected context, read The Anticipatory Duet: A Reactive Movement Partner Conditioned on Pre-Movement Intention and Thinking in Movement: Maxine Sheets-Johnstone and the Intent Layer as a Site of Cognition, Not Execution.

References

Pöppel, E. (1997). A hierarchical model of temporal perception. Trends in Cognitive Sciences, 1(2), 56–61. https://doi.org/10.1016/S1364-6613(97)01008-5

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

Wittmann, M. (2011). Moments in time. Frontiers in Integrative Neuroscience, 5, 66. https://doi.org/10.3389/fnint.2011.00066

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