Researchers have developed FabriVLA, a compact vision-language-action model that achieves 92% success across 50 different robotic manipulation tasks on the Meta-World MT50 benchmark. The model uses a 1-billion parameter vision-language backbone combined with a flow-matching action head, and is trained end-to-end in a single stage. This demonstrates that precise multi-task manipulation doesn't require massive, billion-parameter VLMs.
What the Researchers Built
FabriVLA is a lightweight Vision-Language-Action (VLA) model designed for precise multi-task robotic manipulation. It consists of three main components: an InternVL3.5 vision-language backbone (around 1 billion parameters), a state encoder that processes end-effector pose and object positions, and a flow-matching action head that generates action sequences.
The action head incorporates two novel design elements. Gated self-attention allows action tokens at different time steps to attend to each other through a learnable gate that starts closed and gradually opens during training. This helps the model capture temporal dependencies across the 50-step action horizon without destabilizing early optimization. Shallow VLM layer fusion combines the final VLM layer with an intermediate layer, exposing both high-level semantic context and low-level spatial details to the action head—critical for precise object localization and contact-rich tasks.
FabriVLA is trained in a single stage on the Evo-1 Meta-World dataset (2,500 trajectories across 50 tasks). All parameters, including the VLM backbone, are unfrozen and jointly optimized from the start. The entire model fits on a single GPU and doesn't require the multi-billion-parameter backbones used by many contemporary VLA models.
Key Results
On the Meta-World MT50 benchmark—50 manipulation tasks across four difficulty tiers—FabriVLA achieves a tier-average success rate of 90.0% and an overall episode-level success rate of 92.0%. The model performs consistently across all difficulty levels:
| Difficulty Tier | Success Rate |
|---|---|
| Easy | 95.0% |
| Medium | 88.2% |
| Hard | 86.7% |
| Very Hard | 90.0% |
| Tier Average | 90.0% |
| Overall Episode | 92.0% |
A controlled ablation study (with a frozen VLM backbone for efficiency) reveals that the gated self-attention mechanism is the decisive component. Adding gated self-attention to the base action head raises tier-average success from 48.5% to 57.7% and overall episode success from 55.4% to 66.9%. Shallow VLM layer fusion further contributes in the full model, improving the deep-only baseline by 7.1 and 5.2 percentage points on tier-average and overall scores respectively.
How It Works
FabriVLA treats action generation as a flow-matching problem. Instead of predicting actions directly, it learns a time-dependent vector field that smoothly transforms uniform noise into expert action chunks. The action head predicts a 50-step action horizon (chunk of actions) that is executed in a receding-horizon fashion.
Gated self-attention is implemented as a transformer block where action tokens can attend to each other, but the attention is gated by a learnable scalar initialized to zero. At the start of training, the gate is closed—the block behaves like cross-attention only, where each action token attends only to the VLM context. As training progresses, the gate opens, gradually allowing the model to learn inter-step dependencies. This avoids the optimization difficulties that arise when randomly initialized action tokens are asked to attend to each other from scratch.
Shallow VLM layer fusion concatenates the output of the final VLM layer with an intermediate layer (layer 6 of InternVL3.5). The final layer provides rich semantic understanding for task-level reasoning, while the intermediate layer retains lower-level spatial features needed for precise positioning and contact. This fused representation is fed into the action head’s cross-attention mechanism.
The state encoder uses a two-layer MLP with ReLU activation to embed the robot state (end-effector pose, gripper state, object positions) into a single token that is prepended to the VLM context sequence. Training runs for 150k steps on a single GPU with mixed precision and DeepSpeed ZeRO-2 optimization.
Why This Matters for Robotics
FabriVLA shows that a VLA model with only a 1B-parameter backbone can match or exceed the performance of much larger models on multi-task manipulation benchmarks. This has direct implications for cost and deployment. Smaller models require less GPU memory, lower inference latency, and can run on edge hardware—making them practical for real-world robot applications.
The architecture is modular: the flow-matching action head, gated self-attention, and layer fusion are all transferable design choices that can be applied to other VLM backbones. This opens the door for lightweight models that can be trained on smaller datasets and fine-tuned for specific tasks without massive computational budgets.
For robot buyers and engineers evaluating used industrial robots or warehouse robots, a compact VLA like FabriVLA could enable multi-task capabilities without requiring the latest high-end GPUs. It also suggests that future generation robots might achieve versatile manipulation with relatively modest onboard compute.
Limitations and Open Questions
All evaluations were performed in simulation (Meta-World); real-world transfer remains to be demonstrated. The benchmark covers only 50 tasks, and while varied, they don't capture the full breadth of industrial manipulation scenarios. The action horizon is fixed at 50 steps, which may not be sufficient for long-horizon tasks requiring hundreds of actions.
The ablation study was conducted with a frozen VLM backbone and reduced training steps, so the relative contribution of gated self-attention in the full jointly-trained model may differ from the reported numbers. Additionally, the model's performance on unseen tasks (zero-shot generalization) was not explored.
Finally, FabriVLA uses a specific VLM backbone (InternVL3.5). It's unclear how well the design choices generalize to other backbones or why layer 6 was chosen for shallow fusion.
Frequently Asked Questions
What makes FabriVLA different from other VLA models? FabriVLA uses a flow-matching action head with gated self-attention and shallow layer fusion, all trained jointly without freezing, achieving high performance with only a 1B parameter VLM backbone.
How does FabriVLA compare to models using larger backbones? FabriVLA matches or exceeds reported results from many larger VLA models on Meta-World MT50, achieving 92% overall success despite using a comparatively small vision-language model.
Can FabriVLA be deployed on real robots? The current work is limited to simulation, but the model's small size (1B parameters) makes it a strong candidate for real-world deployment on edge hardware. Real-robot experiments are a natural next step.
What tasks does FabriVLA handle? It performs 50 tasks from the Meta-World MT50 suite, ranging from simple reaching and pushing to complex tasks like door opening, peg insertion, and nut assembly.
Conclusion
FabriVLA achieves state-of-the-art results on multi-task manipulation with a lightweight architecture that doesn't rely on massive VLMs. Its gated self-attention and shallow layer fusion designs are practical contributions that could influence future VLA model development. The model demonstrates that compact, efficient VLA models can deliver precise, multi-task manipulation performance.