Single-source brief: A research team reports a new framework, but no independent confirmation was supplied.
What changed
The team introduced HAF, short for Humanoid Adaptation Framework.
HAF aims to adapt generalist vision-language-action, or VLA, models for humanoid robots.
The authors say standard VLA models struggle with full-body work. A humanoid must coordinate walking, waist pose, and both arms.
HAF has two parts: HAF-VLA and HAF-Steer.
HAF-VLA breaks action generation into three steps. The authors say this preserves links between body motions.
HAF-Steer refines behavior with reinforcement learning. It keeps the main VLA model frozen during that process.
Why the design matters
The paper frames this as a control problem, not just a language-model problem.
A robot can fail when arm motion conflicts with balance or foot placement. HAF seeks to avoid those mismatched full-body actions.
The team also says direct tuning of large VLA models can need heavy compute. It may also create safety risks during real-robot learning.
HAF-Steer limits learning to a smaller noise space. The authors say this reduces the need to update the full backbone.
Deployment reality
The authors report tests on seven real-world humanoid loco-manipulation tasks. They say HAF beat single-stage VLA baselines in task performance and whole-body coordination.
This is a research result, not a confirmed commercial deployment. The supplied evidence does not name the robot hardware, task details, runtime, payload, or safety limits.
It also provides no independent replication or operating data.