New preprint: Humanoid-DART combines diffusion-based trajectory generation and reinforcement learning to expand humanoid loco-manipulation skills from sparse demonstrations. [paper]