Large robot models can do many tasks but often need fine-tuning to be reliable, and retraining the whole model is expensive.
The authors propose LoRN, which guides the model by changing only the initial random noise it uses, through a small set of directions. This avoids updating the model and works with video noise in world-action models that predict video and actions.
Across tests, LoRN improved task success by 26.2 percentage points, much more than prior methods. This makes it practical to improve large robot policies without costly retraining.
