Embodied AI Glossary中文

Foot Skating / Floating / Penetration

脚滑 / 漂浮 / 穿地(动作数据伪影)Advanced

Three common physically-impossible foot errors that appear when human motion is retargeted onto a robot.

These are the three most common physically implausible artifacts that show up when human motion — from motion capture or video pose estimation — is retargeted (converted into robot joint trajectories). Foot skating is when a foot that should be planted firmly on the ground instead slides horizontally; floating is when a foot that should be touching the ground instead hovers above it; penetration is when a foot sinks below the ground surface. These mostly come from scaling errors caused by differences in height and leg-length proportions between the human and the robot, plus errors in the video pose estimation itself. They're not always obvious to the eye, but when a motion-tracking policy is trained in physics simulation, the robot simply can't reproduce these motions, which hurts balance learning and increases tracking error. The usual fix is to add constraints during retargeting — forcing foot height to match the ground closely on contact frames, damping horizontal foot velocity during contact, or filtering out problematic segments by threshold. Works such as GMR and PHUMA specifically handle these artifacts.

ExamplePHUMA measures this with three metrics: a foot within 1cm of the ground on a contact frame counts as not floating, sinking less than 1cm counts as not penetrating, and horizontal velocity below 10cm/s counts as not skating.

Also called
Motion Retargeting Artifacts, Foot Sliding, Ground Penetration
Related
Motion Retargeting · General Motion Retargeting · PHUMA · Motion Tracking · Motion Capture · Interpenetration
Sources
PHUMA: Physically Reliable Humanoid Locomotion Dataset (arXiv)
Retargeting Matters: General Motion Retargeting for Humanoid Motion Tracking (arXiv)

See it in the full glossary →