MergeVLA: Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent

2025-11-24

AI Summary

VLA models suffer significant performance degradation when merging cross-skill capabilities, primarily due to inter-task parameter interference and strong inter-layer dependencies in action modules. To address this, we propose MergeVLA: (1) a sparse-activation LoRA adapter with task-specific masks to enforce parameter consistency across tasks; (2) a decoupled, pure cross-attention action module that eliminates inter-layer dependencies; and (3) an unsupervised test-time task routing mechanism enabling plug-and-play multi-skill composition. Built upon vision-language foundation models, MergeVLA supports joint multi-task training without additional annotations. Evaluated on LIBERO, LIBERO-Plus, RoboTwin, and a real SO101 robotic arm, MergeVLA matches or surpasses single-task expert models—achieving, for the first time, a VLA agent that is highly generalizable, mergeable, and fine-tuning-free.

Technology Category

• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Computer Vision: Multi-modal Vision
• Machine Learning: Mixture of Experts (MoE)

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Recent Vision-Language-Action (VLA) models reformulate vision-language models by tuning them with millions of robotic demonstrations. While they perform well when fine-tuned for a single embodiment or task family, extending them to multi-skill settings remains challenging: directly merging VLA experts trained on different tasks results in near-zero success rates. This raises a fundamental question: what prevents VLAs from mastering multiple skills within one model? With an empirical decomposition of learnable parameters during VLA fine-tuning, we identify two key sources of non-mergeability: (1) Finetuning drives LoRA adapters in the VLM backbone toward divergent, task-specific directions beyond the capacity of existing merging methods to unify. (2) Action experts develop inter-block dependencies through self-attention feedback, causing task information to spread across layers and preventing modular recombination. To address these challenges, we present MergeVLA, a merging-oriented VLA architecture that preserves mergeability by design. MergeVLA introduces sparsely activated LoRA adapters via task masks to retain consistent parameters and reduce irreconcilable conflicts in the VLM. Its action expert replaces self-attention with cross-attention-only blocks to keep specialization localized and composable. When the task is unknown, it uses a test-time task router to adaptively select the appropriate task mask and expert head from the initial observation, enabling unsupervised task inference. Across LIBERO, LIBERO-Plus, RoboTwin, and multi-task experiments on the real SO101 robotic arm, MergeVLA achieves performance comparable to or even exceeding individually finetuned experts, demonstrating robust generalization across tasks, embodiments, and environments.

Problem

• VLA models fail when merging experts trained on different robotic tasks
• LoRA adapters diverge beyond capacity of existing merging methods
• Action experts develop dependencies preventing modular recombination

Innovation

• Sparsely activated LoRA adapters via task masks
• Cross-attention-only blocks for action expert
• Test-time task router for unsupervised task inference