By decomposing humanoid control into hierarchical stages and constraining RL to a learned latent space, you can adapt large VLA models to real robots efficiently without retraining the backbone or excessive real-world exploration.
This paper presents HAF, a framework for adapting general-purpose vision-language-action models to humanoid robots performing complex whole-body tasks. The approach uses hierarchical action generation to coordinate locomotion, posture, and arm manipulation, then refines the policy through efficient reinforcement learning in a compact latent space without updating the large foundation model.