Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.25802 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917531614183424 |
|---|---|
| author | Lin, Weifeng Huang, Siyuan Li, Hao Chen, Tingwei An, Ruichuan Wei, Xinyu Liu, Jianbo Li, Hongsheng |
| author_facet | Lin, Weifeng Huang, Siyuan Li, Hao Chen, Tingwei An, Ruichuan Wei, Xinyu Liu, Jianbo Li, Hongsheng |
| contents | Vision-Language-Action (VLA) models widely adopt pretrained Vision-Language Models (VLMs) as policy backbones, yet it remains unclear what kind of pretrained VLM representation is useful as a VLA initialization. In this paper, we study VLA initialization as a controlled representation-design problem along three axes: capability-level embodied VQA supervision, parameter-update strategy, and robot-data pretraining. Our experiments show that the original pretrained VLM representation is a key source of action performance. However, embodied VQA adaptation does not yield uniform gains: its benefit depends on downstream bottlenecks, and gains from different capability domains are not simply additive. For update strategy, LoRA provides a more reliable initialization than Full Finetune, indicating that overly reshaping the pretrained representation can weaken VLA initialization. Robot-data pretraining further improves VLA initialization, with the strongest variant obtained by staged LoRA-based training. Together, these findings suggest that effective VLM-to-VLA adaptation should inject action-relevant embodied and robot-trajectory signals while preserving the pretrained VLM representation that remains useful for action learning. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_25802 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Rethinking VLM Representation for VLA Initialization Lin, Weifeng Huang, Siyuan Li, Hao Chen, Tingwei An, Ruichuan Wei, Xinyu Liu, Jianbo Li, Hongsheng Computer Vision and Pattern Recognition Vision-Language-Action (VLA) models widely adopt pretrained Vision-Language Models (VLMs) as policy backbones, yet it remains unclear what kind of pretrained VLM representation is useful as a VLA initialization. In this paper, we study VLA initialization as a controlled representation-design problem along three axes: capability-level embodied VQA supervision, parameter-update strategy, and robot-data pretraining. Our experiments show that the original pretrained VLM representation is a key source of action performance. However, embodied VQA adaptation does not yield uniform gains: its benefit depends on downstream bottlenecks, and gains from different capability domains are not simply additive. For update strategy, LoRA provides a more reliable initialization than Full Finetune, indicating that overly reshaping the pretrained representation can weaken VLA initialization. Robot-data pretraining further improves VLA initialization, with the strongest variant obtained by staged LoRA-based training. Together, these findings suggest that effective VLM-to-VLA adaptation should inject action-relevant embodied and robot-trajectory signals while preserving the pretrained VLM representation that remains useful for action learning. |
| title | Rethinking VLM Representation for VLA Initialization |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2605.25802 |