A 1-Billion-Parameter Robot AI Topped a 50-Task Manipulation Benchmark

A 1-billion-parameter model called FabriVLA, built by Chinese robotics lab Youibot (优艾智合), has taken the top spot on the Evo-SOTA leaderboard for Meta-World MT50 (a 50-task robot-manipulation suite used across the field) at 90.0% tier-average success. The larger models it displaced, including the well-known π0, got there by adding parameters. FabriVLA got there by redesigning two pieces of the network.

The change matters because Meta-World MT50 is not a chat benchmark. It is a 50-task test of multi-step pick, place, and insert manipulation, the kind that real factory work demands. For two years, the scoreboard has rewarded the largest vision-language-action (VLA) models, networks that read a camera image, parse a language instruction, and emit a robot action sequence. The standard assumption was that more parameters would keep buying more competence. FabriVLA, at 1 billion parameters, sits at the top of the same leaderboard.

The two moves are unglamorous and specific. The first is gated self-attention applied across the action tokens: each predicted step in a robot's motion plan is allowed to look at the steps before and after it, so a long pick-then-insert sequence stays coherent instead of drifting. Standard action heads process each step independently. On a long horizon, the pick can end aligned but the insert can land two centimeters off-center. The second is a fusion of deep and shallow visual features. Deep features tell the robot what an object is; shallow features tell it where the gripper should land. Most VLA models pick one layer. FabriVLA uses both.

The ablation data is what makes the case. Pull out the deep-plus-shallow visual fusion and the tier-average drops from 90.0% to 82.9%, a 7.1-point swing. Pull out the gated self-attention inside the action head and the model loses its largest single contribution. The two changes compound. Neither is a parameter increase. FabriVLA also skips the usual large-scale robot-data pretraining and trains end-to-end in a single stage.

The regime change this signals is what makes the scoreboard move worth more than a single benchmark flip. Embodied AI has been running a parameter race, the same way language models did from 2020 to 2023. The argument from the QbitAI report and the paper itself is that the embodied version of that race is closing: the right network shape beats raw scale on the multi-task manipulation that actually matters on a factory line, and the smaller footprint is a feature, not a bug. A 1-billion-parameter VLA can run on the robot's own compute instead of a cloud round-trip, which means lower latency and no network dependency.

Youibot's industrial context gives the result weight. The company has been deploying mobile robots in semiconductor fabs, energy and chemical plants, and lithium-battery lines for nine years, and says it has more than 800 live scenarios in production. That is not a benchmark pedigree. It is a constraint: factory work punishes latency, bandwidth, and unreliable connectivity in ways that a cloud-only VLA cannot absorb. A 1-billion-parameter model that runs on-device is the answer to a problem the company already has.

There is a longer commercial story the source wants told: companion launches, a humanoid called 隙锋, a stated goal of 10,000 industrial sites in three years. That material appears in the QbitAI report and reads as company positioning rather than verified deployment. The benchmark and the paper are the load-bearing claims here. The scoreboard moved, and the mechanism that moved it is reproducible from the ablation table.

The Evo-SOTA leaderboard recomputes as new submissions land, and the tier-average framing that puts FabriVLA at 90.0% is one slice of a richer scorecard. Raw success rate and hardest-tier breakdowns can rank models differently. The architecture-race claim survives either way. The specific ordering may not.