The field of humanoid robot control is moving towards more versatile and naturalistic control methods. Recent developments focus on learning skills from human motions and synthesizing novel ones, enabling zero-shot task-specific control and bridging the sim-to-real gap. A key direction is the integration of advanced physics models and control architectures to achieve highly dynamic motions and acrobatic maneuvers. Another area of progress is the development of unified frameworks for whole-body humanoid imitation, allowing for the creation of generalized humanoid controllers that can be applied to various robot morphologies. Noteworthy papers include:

• BeyondMimic, which introduces a guided diffusion framework for learning from human motions and enables zero-shot task-specific control.
• GBC, which establishes a comprehensive pathway from human motion to robot action through a unified framework.
• MASH, which proposes a cooperative-heterogeneous multi-agent reinforcement learning method for single humanoid robot locomotion, accelerating training convergence and improving whole-body cooperation ability.