[2512.08333] Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging

Computer Science > Robotics arXiv:2512.08333 (cs) [Submitted on 9 Dec 2025 (v1), last revised 27 Feb 2026 (this version, v3)] Title:Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging Authors:Yajat Yadav, Zhiyuan Zhou, Andrew Wagenmaker, Karl Pertsch, Sergey Levine View a PDF of the paper titled Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging, by Yajat Yadav and 4 other authors View PDF HTML (experimental) Abstract:Generalist robot policies, trained on large and diverse datasets, have demonstrated the ability to generalize across a wide spectrum of behaviors, enabling a single policy to act in varied real-world environments. However, they still fall short on new tasks not covered in the training data. When finetuned on limited demonstrations of a new task, these policies often overfit to the specific demonstrations--not only losing their prior abilities to solve a wide variety of generalist tasks but also failing to generalize within the new task itself. In this work, we aim to develop a method that preserves the generalization capabilities of the generalist policy during finetuning, allowing a single policy to robustly incorporate a new skill into its repertoire. Our goal is a single policy that both learns to generalize to variations of the new task and retains the broad competencies gained from pretraining. We show that this can be achieved through a simple yet effective strategy: interpolating the weights of...