SMR: Learning streaming motion retargeting for humanoids from human semantic

Key points

• Prior physics-supervised neural methods train bidirectional full-sequence models on physics-refined kinematic trajectories, remaining tied to prescribed kinematic solutions and requiring noncausal context.1
• We propose SMR, a streaming neural motion retargeter learned from semantic-conditioned physics experts.1
• We convert human motion into robot-aware task-space targets for body landmark positions, orientations, and velocities.1
• Experiments demonstrate that semantic-conditioned physics experts generate motion with higher semantic fidelity and physical validity than direct optimization-based retargeting, while the resulting streaming retargeter achieves improved closed-loop performance over kinematic and physics-refined baselines under a common whole-body controller.1

Sources (1)

• [1]SMR: Learning streaming motion retargeting for humanoids from human semantic
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 03:59 AM

  Retargeting human motion to humanoids requires preserving motion semantics while accounting for the robot's physical constraints.

  Prior physics-supervised neural methods train bidirectional full-sequence models on physics-refined kinematic trajectories, remaining tied to prescribed kinematic solutions and requiring noncausal context.

Extractive summary: sentences quoted from the sources.