HAF-VLA: Hierarchical Action Flow Framework for Humanoid Vision-Language-Action Control
A humanoid robot control model that plans walking first, then the waist, then the arms, scored 68.3% on six real household and factory tasks against 53.3% for Physical Intelligence's pi0.5, according to this preprint from Peking University and the Beijing Innovation Center of Humanoid Robotics. The method reuses pi0.5's pretrained weights rather than training a new model.
Can a vision-language-action model built for fixed or wheeled robots be adapted to a full-size walking humanoid by generating its whole-body actions in a fixed order, legs and head first, then waist, then arms and hands?
Humanoids must coordinate walking, posture, gaze and two arms at once. When a single model generates all 26 control dimensions in one pass, small errors in base or waist motion force the upper body to compensate, which hurts precision and stability. Most VLA models were designed for tabletop or wheeled manipulators and do not respect this chain of dependencies.
Generalist flow-matching VLAs such as pi0.5 denoise the full action chunk in one stage. Humanoid foundation models such as NVIDIA's GR00T N1.7 and world-model policies such as Cosmos Policy train for whole-body control, and classic imitation learning such as ACT learns one task at a time. None of them orders the action generation by body part.
HAF-VLA starts from pi0.5 and keeps its vision-language encoder and action expert. It splits generation into three cumulative stages inside one shared 32-dimension action space (26 used): locomotion and head, then waist, then arms and hands, with 10 denoising steps each. After each stage, the clean prediction is re-encoded into a key-value cache that conditions the next stage, so later stages can refine earlier motion rather than overwrite it. Training used 120 teleoperated demonstrations per task on the Tian Gong 3.0 and 2.0 humanoids.
Over 10 real-robot rollouts per task, HAF-VLA averaged 68.3% normalized score across six tasks, against 53.3% for pi0.5, 37.8% for GR00T N1.7, 25.6% for ACT and 25.0% for Cosmos Policy. It scored 93.3% on Box Transfer (pi0.5 60.0%), 56.7% on Ball Tossing (33.3%) and 66.7% on Laundry Loading (53.3%), and tied pi0.5 at 53.3% on Clothes Retrieval. In ablations on a Box Carry task, the legs-waist-arms order scored 86.7% versus 70.0% waist-first and 60.0% arms-first; pi0.5 with 30 denoising steps scored 30.0%, so extra steps alone do not explain the gain. Removing the cache only at deployment produced unsafe motion. With the target table moved 30 cm, HAF-VLA kept 73.3% against 40.0% for pi0.5.
The authors note that three stages raise the total denoising budget and latency (0.123 s per chunk versus 0.075 s for 10-step pi0.5), and that the method depends on the base model's priors. We add that all tests ran on two robots from one lab with 10 rollouts per setting, so a single rollout can move a score by several points; Cosmos Policy's weak score partly reflects latency in the authors' implementation. It is a preprint and has not been peer reviewed.
The result matters to humanoid makers that license or fine-tune general VLA models instead of training their own, including users of Physical Intelligence's pi0.5 and NVIDIA's GR00T. A staged action order is cheap to add to an existing model, so Chinese humanoid programmes around the Beijing centre could adopt it within months; the latency cost is the main barrier for faster tasks.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.