Humanoid robots today often need a separate controller for each task, like walking, recovering from pushes, or picking up objects. This wastes effort because each controller relearns balance and coordination from scratch.
Metatron solves this by training one motion model that accepts partial instructions called prompts. After training, the model is frozen and never changed. New tasks are added by writing ordinary code that sends prompts at each control step, guiding the model's movements.
On a real robot, one frozen model handles walking, recovering, picking up objects, and sitting. It succeeds more often at reaching and handles pushes better than systems that follow exact motion references. This means new skills can be added just by writing code, without retraining the model.