Give StepControl only when each foot lands, and full-body motion follows through long rollouts
Synopsis
Human motion generation has mostly been steered by naming an action or a place to move; StepControl instead takes only a future left/right foot-contact schedule, with no target root or body poses and no semantic prompt. It treats contacts as explicit temporal events, injects them through dedicated cross-attention with classifier-free guidance, and adds a stochastic mixture noise schedule and warmup for stable long-horizon autoregressive generation. Experiments report that it follows prescribed contact timing more accurately than adapting existing spatially guided generators to contact constraints, keeps contact control over long continuous rollouts, and responds to edited schedules such as cadence changes, stop-and-go pauses, limping gaits and bilateral jumping.
Interpretation
Full-body motion can be driven by a future left/right foot-contact schedule rather than by stating an action or a spatial target. Controllable human motion generation has largely focused on specifying what action should occur or where the body should move, leaving contact timing outside the control interface. Against adapting existing spatially guided motion generators to contact constraints, StepControl follows the prescribed contact timing more accurately.
Contacts are represented as explicit temporal events and injected through dedicated cross-attention, combined with classifier-free guidance. The control signal is a set of discrete landing moments rather than pose targets or text prompts. A stochastic mixture noise schedule and a warmup strategy are introduced for stable long-horizon autoregressive generation.
Contact control remains effective over long continuous rollouts and tracks edited schedules. The edits it responds to include cadence changes, stop-and-go pauses, limping gaits and bilateral jumping. These behaviors are shown in long continuous rollouts.
Perspective
The demonstrated setting is human motion generation conditioned on a future left/right foot-contact schedule, maintaining contact control over long continuous autoregressive rollouts and responding to edited schedules such as cadence changes, pauses, limping gaits and bilateral jumping.
The reported comparison is more accurate contact-timing following, without specific error values; long-horizon effectiveness rests on behavior in continuous rollouts, and schedule editing is limited to the listed kinds of change.