HAF: Adapting Generalist VLAs to Humanoid Whole-Body Loco-manipulation via Hierarchical Action Flow and Spectral Latent RL
HAF keeps the large VLA frozen and adapts humanoid control through staged action generation plus compact latent RL.
The paper says humanoid loco-manipulation breaks down when a single-stage VLA tries to produce locomotion, waist posture, and dual-arm actions all at once. HAF-VLA splits full-body action denoising into three sequential stages while preserving cross-stage kinematic dependencies. HAF-Steer then refines behavior in a reduced noise subspace with regularized SAC, avoiding direct updates to the VLA backbone. The authors report gains over vanilla single-stage VLA baselines on seven real-world humanoid tasks. ArXiv · AI/CL/LG's note
The paper says humanoid loco-manipulation breaks down when a single-stage VLA tries to produce locomotion, waist posture, and dual-arm actions all at once. HAF-VLA splits full-body action denoising into three sequential stages while preserving cross-stage kinematic dependencies. HAF-Steer then refines behavior in a reduced noise subspace with regularized SAC, avoiding direct updates to the VLA backbone. The authors report gains over vanilla single-stage VLA baselines on seven real-world humanoid tasks. ArXiv · AI/CL/LG's note
score 5