FF EAI says its robot AI took first place on RoboColiseum’s simulation leaderboard

FF EAI Robotics Ecosystem says its OmniMind robot AI reached first place on RoboColiseum’s global leaderboard as of November 6, 2026. The company reported a score of 0.8698 out of 1 in the simulated evaluation, presenting the result as evidence of progress in the software that interprets instructions and directs robot movement.

The result was disclosed in a November 8 business update from founder and CEO YT Jia, published through PR Newswire. It is a company-reported benchmark result: the release does not include independent verification of the performance claims. RoboColiseum tests embodied AI—AI that controls a machine’s actions—in simulation rather than through this announcement’s real-world robot trials.

Jia highlighted three component scores alongside the overall result. They cover following instructions, reasoning about space and remaining robust when disturbances occur. FF EAI reported its highest figure among those three in instruction following, with spatial reasoning the lowest.

About 0.99Instruction following

FF EAI reported approximately 0.99 for instruction following.

0.92Spatial reasoning

FF EAI reported 0.92 for spatial reasoning.

0.96Robustness to disturbances

FF EAI reported 0.96 for robustness to disturbances.

Jia attributed OmniMind’s performance to a layered architecture and programmatic execution: carrying out a task through distinct software-controlled steps. He contrasted that design with a typical vision-language-action approach, in which a single model generates actions directly. FF EAI’s description instead assigns separate jobs to instruction interpretation, perception and movement planning.

1. Interpret the request. The model works out the instructions before the later stages determine where and how to move.
2. Locate the target. A 3D perception stage identifies the target’s position, separating that job from interpreting the instruction.
3. Generate a path. Motion planning determines the route the robot should take to carry out the task.

The practical advantage, according to Jia, is that the team can check, analyze and optimize each step separately as the robot performs a task. That is the company’s explanation for its design choice, rather than a separate experimental finding showing which stage produced the leaderboard advantage.

Jia said the team would keep using evaluation feedback to improve OmniMind and its robot skills. He also described real-world deployment as part of the continuing development effort. Those are the next steps he set out; the numerical result announced here concerns performance inside the simulator.