A timetable saying when each foot touches the floor can guide an AI to generate full-body movement, but it cannot specify how that movement feels or how weight passes through the body. That is the useful distinction raised by StepControl, a motion-generation paper first submitted on 28 October 2026. Its sparse instruction is a reminder that controlling an important event is different from controlling every detail of a movement.

Imagine a score with only two lines

Picture a simple animation exercise. One horizontal line represents the left foot and another the right. Dark stretches mean contact with the ground; gaps mean no contact. You draw an alternating pattern, then a short interval with both lines dark, followed by another alternating pattern. There are no arm poses, facial expressions or instructions about mood.

This is a contact schedule: a description of when contact should occur. It is not a set of footprints, which would say where the feet land. Nor is it a force recording, which could describe how strongly a foot pushes. Those are different kinds of information, and keeping them separate helps explain both the attraction and the limits of this interface.

For our imagined animator, the attraction is straightforward. If the important event is a pause before a character continues, drawing that interval may be easier than describing it in a long prompt. The rest of the phrase remains open to generation. This is a hypothetical use case, not a report that the paper tested a particular animation production.

What the new research actually claims

Jang and colleagues' StepControl generates motion in successive segments, conditioned on a future left/right contact schedule. The authors report that their model follows contact timing better than the spatially guided alternatives they adapted for comparison. They also show responses to edited schedules, including cadence changes and pauses (Jang et al., 2026).

The paper therefore supports a bounded claim: contact events can be a useful input for generating movement. It does not show that a two-line score uniquely determines a dance, or that the model senses the floor. The system uses learned movement patterns to supply details that its input leaves unspecified.

There is also an initial movement context: the technical formulation conditions future generation on an observed motion prefix. “From a timetable” describes the control over what follows, not a model creating every detail from two signals with no history.

A score is a choice about what matters

Return to the imagined animation. The same dark interval on both lines could accompany an upright pause, a deeper bend or an upper-body gesture. The two-line notation does not decide among those possibilities. If the character's torso must remain upright, the animator needs another instruction or a separate check.

This is not necessarily a weakness. A creative score often becomes useful by leaving something undecided. A rule can organise attention while allowing variation elsewhere. The important question is whether the interface leaves open the details you want to explore, rather than details you need to protect.

The Lab's earlier discussion of why movement can have more than one valid answer explains this distinction through editing. Here the question is narrower: which temporal events should be fixed before an entire phrase is generated?

Touching the floor does not tell us how weight is organised

A contact flag is a yes-or-no description. Consider two hypothetical frames with both feet on the floor. In one, the performer could be resting mostly over the left leg; in another, the weight could be moving toward the right. The contact flags alone cannot distinguish them because neither foot has to leave the ground for that change to happen.

Likewise, a visible pause does not tell us whether the person is preparing, hesitating or finishing. An observer may interpret those qualities, but the contact schedule does not encode an explanation. Adding them to the story without another source would confuse interpretation with measurement.

For somatic practice, this is the crucial boundary. Attention to support, pressure and readiness can be part of a mover's experience. A model's binary floor-contact input is a deliberately reduced description of an external relation. It can help coordinate a generated body without reproducing the experience of having one.

Three checks for our imaginary animation

Suppose the animator now has several generated versions. A useful review would separate three questions.

Did the scheduled events happen? The left and right contacts should occur in the requested intervals. This tests the control input on its own terms. It does not yet say whether the whole motion is useful.

Does the rest of the phrase make sense? The animator can inspect transitions, unwanted foot sliding and the relation between the lower and upper body. A convincing contact timetable could coexist with an awkward arm movement. A single success label would conceal that difference.

Does the version serve the intended scene? A pause intended as uncertainty could read as confidence. That judgement concerns staging and interpretation, not just feet. The person making the scene remains responsible for deciding whether the generated choice works.

These are proposed review questions, not additional experiments reported by StepControl. Their purpose is to prevent one successful control channel from standing in for a complete assessment.

Why this matters beyond a new model

Generative interfaces often invite us to describe everything in words. A contact timeline suggests a different design possibility: let people specify the relation they care about in a form where it can be inspected. A timing strip may communicate a pause more precisely than an adjective, while a separate visual control may be better for direction or reach.

That does not make words obsolete. It gives each input a clearer job. Language might establish the broad action, a timeline its events, and a human review the expressive result. This is a design proposal inspired by the paper, not a claim about an implemented combined interface.

The same separation should extend to records. If contact was prescribed, inferred from video or measured by a sensor, those origins should remain distinguishable. The Lab's guide to recording observed, inferred and generated evidence shows why the distinction matters when a system behaves unexpectedly.

So can two lines choreograph a body? They can constrain a useful part of its timing while a model proposes the rest. The more honest ambition is not to compress the whole person into a footfall chart. It is to give the person making the work a small, intelligible control—and a clear view of everything that control leaves undecided.

References

  • Jang, H., Bae, J., Lim, D., Kim, T., Jeong, S., Lee, Y., Kim, Y. S., & Kim, Y. M. (2026). StepControl: Long-horizon human motion generation from foot-contact schedules. arXiv. https://arxiv.org/abs/2610.38153