Surpassing Pi-Zero: Youibot FabriVLA 1B-Parameter Model Topping the Embodied Intelligence Leaderboard Marks the Shift From Parameter Competition to Architecture Competition

Youibot FabriVLA with 1B parameters topped Evo-SOTA Meta-World MT50 at 90.0% tier-average, surpassing Physical Intelligence Pi-Zero. The achievement is significant because MT50 tests simultaneous mastery of 50 manipulation tasks, precisely the multi-task capability industrial applications require.

The core innovation is a redesigned neural architecture specifically optimized for robot manipulation rather than following the dominant trend of blind parameter scaling. FabriVLA introduces two key architectural innovations: deep and shallow visual feature fusion that enables the model to simultaneously understand semantic task objectives through high-level vision and precise spatial positioning through low-level visual details, and a gated self-attention mechanism that ensures each step in the action sequence references preceding and following steps, learning smooth action transitions and coordination. Ablation studies confirm the contributions: removing deep-shallow fusion drops tier-average success rate from 90.

0% to 82.9%, while gated self-attention is identified as the highest-contributing component in the action head. The 1B parameter achievement demonstrates that architecture optimization specifically tuned to robot manipulation can outperform brute-force parameter scaling.

Youibot has 800+ real-world industrial deployments across semiconductor, energy, and battery sectors. The Fengxi humanoid robot received 4,000 unit orders on launch day. The leaderboard ranking validates the 1B model as production-ready for edge-deployed industrial manipulation.

The smaller parameter count has direct commercial implications. Larger models require more computing power, higher latency, and often cloud dependency, limiting real-time industrial deployment. A 1B-parameter model that achieves state-of-the-art manipulation performance can run on edge computing hardware inside the robot, eliminating cloud dependency, reducing latency, lowering deployment cost, and improving reliability in