| name | fabrivia-lightweight-vision-language-action-multi-task-manipulation |
| description | FabriVLA: lightweight VLA model combining InternVL3.5 VLM backbone with flow-matching action head. Gated self-attention across action tokens, shallow VLM layer fusion. 90.0% success on Meta-World MT50. 1B-scale VLM without billion-parameter backbone. Activation: vision-language-action, lightweight model, multi-task manipulation, flow-matching, robot learning. |
| metadata | {"arxiv_id":"2607.08575","published":"2026-07-09","authors":"Shiyuan Yang, Borong Zhang, Jizheng Zhang, Zhijia Tao, Junfei Guo","tags":["vision-language-action","lightweight-model","multi-task-manipulation","flow-matching","robot-learning"]} |
FabriVLA: A Lightweight Vision-Language-Action Model for Precise Multi-Task Manipulation
Overview
FabriVLA is a lightweight Vision-Language-Action model that demonstrates a compact VLA built on a 1B-scale VLM can achieve strong performance without relying on multi-billion-parameter VLA backbones. It combines an InternVL3.5 vision-language backbone with a flow-matching action head, trained via single-stage joint optimization.
Key Innovations
Compact Architecture
- InternVL3.5 vision-language backbone (1B scale)
- Flow-matching action head with gated self-attention across action tokens
- Shallow VLM layer fusion for enriched spatial context
- No reliance on multi-billion-parameter backbones
Single-Stage Joint Optimization
- Trained from pretrained VLM and randomly initialized action head
- Joint optimization in a single stage
- Avoids complex multi-stage training pipelines
Meta-World MT50 Performance
- 90.0% tier-average success rate across 50 diverse manipulation tasks
- Demonstrates that compact VLA models can match larger architectures
- Strong performance across diverse task types
Methodology
- Backbone Selection: InternVL3.5 as VLM backbone (1B scale)
- Action Head Design: Flow-matching head with gated self-attention
- Fusion Strategy: Shallow VLM layer fusion for spatial context
- Training: Single-stage joint optimization from pretrained VLM + random action head
- Evaluation: Meta-World MT50 benchmark (50 tasks)
Implications
- Lightweight VLA models are viable alternatives to large-scale VLA architectures
- 1B-scale VLMs provide sufficient vision-language understanding for manipulation
- Flow-matching action heads with gated attention are effective for action generation
- Single-stage training simplifies VLA deployment
Pitfalls
- Meta-World MT50 is simulation-only — real-world performance not demonstrated
- 1B-scale VLM may struggle with complex language instructions
- Single-stage training may not be optimal for all task domains
- Flow-matching may be slower than direct action prediction at inference
Activation Keywords
vision-language-action, FabriVLA, lightweight VLA, multi-task manipulation, flow-matching, InternVL3.5, robot learning, Meta-World MT50, compact model
Paper Reference
arXiv:2607.08575 - "FabriVLA: A Lightweight Vision-Language-Action Model for Precise Multi-Task Manipulation" (Jul 2026)