Robot Motion Transition with Human Motion Prior

How do you join two humanoid motions that were generated or recorded independently, so that a tracking policy can actually execute the combined sequence on a robot?

Humanoid motion libraries are built clip by clip. When two clips are chained, nobody has said where the second one should begin relative to the first, which way the robot should face, or how long the transition should take. Two clips can each be fine on their own and still be hard to connect under an arbitrary placement, which shows up as foot sliding, extra turning or unwanted steps.

The common fix is to snap the second clip's start to the first clip's end and interpolate the root position and heading. Generative in-betweening models such as Kimodo or MotionBricks produce richer motion but keep the original boundary arrangement, so a badly placed clip makes the generator invent unintended locomotion.

STAMP (Spatio-Temporal Alignment with a Motion Prior) works in human motion space. Both robot clips are retargeted to a human rig; a learned alignment model, two small Transformer encoders, scores 388 candidate placements of the second clip (planar offsets, headings from minus 60 to 90 degrees, durations of 0.5 to 2.0 s) and picks the cheapest. A diffusion-based human motion prior then generates the transition under that placement, and the finished sequence is retargeted back to the robot and tracked by a BeyondMimic policy. Because both models live in human space, they are reused across robot bodies.

On 18 motion pairs from boxing, industrial maintenance and traffic-control scenarios, each evaluated over 1,024 simulated rollouts: Unitree G1 success 80.69% for STAMP vs 69.59% for root and heading interpolation, 56.91% for MotionBricks and 50.69% for the Kimodo in-betweener without alignment; body velocity error 0.229 vs 0.258 m/s for interpolation. ROBOTIS K1 success 87.52% vs 84.70% interpolation and 64.55% Kimodo. Picking the rank-1 rather than the rank-100 alignment candidate added 9.36 points of success on G1 and 15.62 points on K1. On a physical K1 the method produced a connecting step in a running-to-punching transition where interpolation caused foot sliding, and chained four traffic-control motions into longer sequences.

A preprint. Success is measured in simulation against each method's own generated reference, not against a ground-truth transition, and the authors use one training seed per reference. Each composed sequence still needs its own tracking policy, so new combinations are not executed zero-shot. Physical-robot evidence is qualitative plus a five-run motor-current comparison on one motion pair, where the authors note the currents already differed at onset. The K1 gain over interpolation is small (2.8 points).