Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging

Researchers developed a parameter merging technique that allows robot AI policies to learn new tasks while preserving their existing generalist capabilities. The method interpolates weights between finetuned and pretrained models, preventing overfitting and enabling lifelong learning in robotics applications.

• →Generalist robot policies often overfit when learning new tasks, losing their broad capabilities and failing to generalize.
• →Parameter merging by interpolating weights between finetuned and pretrained models preserves generalist abilities while learning new skills.
• →The merged models outperform both pretrained and finetuned models on out-of-distribution variations of new tasks.
• →Performance scaling improves with larger amounts of pretraining data, enhancing the effectiveness of the approach.
• →The method enables continual skill acquisition in lifelong learning scenarios without sacrificing previously learned abilities.