FabriVLA: A Lightweight Vision-Language-Action Model for Precise Multi-Task Manipulation
This work provides a lightweight, efficient VLA model for precise multi-task manipulation, enabling deployment on resource-constrained robotic systems.
FabriVLA achieves a 90.0% tier-average success rate on the Meta-World MT50 benchmark, demonstrating that a compact 1B-scale VLM can deliver strong multi-task manipulation performance without relying on multi-billion parameter backbones.
We present FabriVLA, a lightweight Vision-Language-Action model for Precise Multi-Task Manipulation. FabriVLA combines an InternVL3.5 vision-language backbone with a flow-matching action head featuring gated self-attention across action tokens and shallow VLM layer fusion for enriched spatial context. The model is trained via single stage joint optimization from a pretrained VLM and randomly initialized action head. On the Meta-World MT50 benchmark spanning 50 diverse manipulation tasks, FabriVLA achieves a tier-average success rate of 90.0%, demonstrating that a compact VLA built on a 1B scale VLM can achieve strong performance without relying on multi billion parameter VLA backbones.