Researchers led by Yanjun Li introduced LANGUAGE‑FORCING, a way to train robot world‑action models using language descriptions of task progress. The method aligns text embeddings from a frozen vision‑language model with the robot’s predicted future visuals. It only adds overhead during training, so the robot runs as fast as before. On the HumanoidArena benchmark it reached a 79.17% success rate, beating the next best policy by 9.46 points. In five real‑world tasks the success rose from 53.33% to 71.33%.
Why it matters
For developers of humanoid robots, the technique offers higher task reliability without slowing down the robot during operation.
