metadata
license: mit
language:
- en
base_model:
- Qwen/Qwen2.5-0.5B
Model Card for MergeVLA-LIBERO
MergeVLA — Single-Skill Experts for Spatial / Object / Goal / Long-10 (LIBERO Task Suite). These models are used as the base expert checkpoints for our MergeVLA.
Model Details
Each uploaded model is a 0.68B-parameter VLA model (excluding the vision backbone) composed of:
- Qwen2.5-0.5B as the Vision-Language Model (VLM)
- A lightweight 0.18B Action Expert
- A two-layer Proprioceptive Projector MLP
✔️ Performance (Success Rates on LIBERO)
| Task Family | Success Rate (%) |
|---|---|
| Spatial | 98.0 |
| Object | 98.6 |
| Goal | 95.0 |
| Long-10 | 95.0 |
🧠 Training Details
Each expert is fine-tuned independently using modified LIBER demonstrations in RLDS format.
| Category | Value |
|---|---|
| LoRA | Enabled (rank = 64) |
| Optimizer | AdamW |
| Learning Rate | 2e-4 |
| Batch Size | 8 (×2 grad accumulation) |
| num_images_in_input | 2 |
Training Steps
- Spatial — 30,000
- Object — 20,000
- Goal — 30,000
- Long-10 — 50,000
Citation instructions
@misc{fu2025mergevla,
title={MergeVLA: Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent},
author={Yuxia Fu and Zhizhen Zhang and Yuqi Zhang and Zijian Wang and Zi Huang and Yadan Luo},
year={2025},
eprint={2511.18810},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2511.18810},
}