Duke University
“CoDance demonstrates that a humanoid robot can learn a complex, physically interactive behavior—partnered dancing—using only a single 9.7-second video of two human dancers.”
Source→“The framework was validated on a physical Unitree G1 humanoid, achieving sustained two-hand dancing with a human partner.”
Source→“The decoupled policy uses an adversarial motion reward [3]. Its discriminator compares two-frame transitions from policy rollouts with transitions from the training references using lower-body and torso poses in the pelvis frame.”
Source→“Our augmentation builds on the quasi-static compliance solve of SoftMimic [4].”
Source→“The robot maintained a consistent following distance (mean of 0.48m to 0.50m against a 0.5m target) and successfully transitioned between forward and backward motions for up to 38.9 seconds without falling”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.