FabriVLA: A Lightweight Vision-Language-Action Model for Precise Multi-Task Manipulation

2026-07-09

AI Summary

This work addresses the challenge of achieving efficient and accurate multi-task robotic manipulation without relying on extremely large models. The authors propose a lightweight vision-language-action (VLA) model with only 1 billion parameters that unifies control across 50 distinct manipulation tasks. Built upon the InternVL3.5 vision-language backbone, the model incorporates a flow-matching action head, gated self-attention mechanisms, and shallow feature fusion, all trained jointly in a single stage. This architecture substantially enhances spatial understanding and action generation fidelity, achieving an average task success rate of 90.0% on the Meta-World MT50 benchmark. The results demonstrate that compact VLA models can attain high performance and remain competitive with significantly larger counterparts.

Abstract

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.

Problem

• Vision-Language-Action
• Multi-Task Manipulation
• Lightweight Model
• Robotic Control
• Precise Manipulation

Innovation

• Vision-Language-Action
• flow-matching
• gated self-attention
• shallow layer fusion
• lightweight VLA