May the Dance be with You: Dance Generation Framework for Non-Humanoids

The paper presents a dance generation framework for non-humanoid agents, teaching them to dance by learning from human videos. The framework combines a reward model using contrastive learning to align optical flow and music features, and reinforcement learning for motion generation. Results show that the agent can create dances that match the music's beat, with user studies preferring the framework over baselines. The study stands out for enabling non-humanoid agents to learn dance without manual design or extensive data collection, and differentiates from previous work on human-like agents and limited real-world robots. Future work includes improving 3D representation and developing a structured motion library for better dance generation.