At the 2026 BAAI Zhiyuan Conference in Beijing, BAAI-incubated startup Xingyuanzhi launched omega-EVA, the world's first embodied interactive world model that redefines how world models participate in robotic decision-making.
From Passive Predictor to Active Feedback: Traditional world models play a bystander role generating training data or offline simulations, remaining outside the real-time control loop. omega-EVA breaks this paradigm with a three-stage Envision-Verify-Act closed-loop architecture.
First, the policy model generates a candidate action proposal based on visual observations and language instructions. Second, this proposal is fed into an action-conditioned latent world model that previews future state changes in compact latent space. Third, a tri-branch refiner module receives the current state, imagined future consequences, and the original proposal to refine the action.
Technical Innovation: omega-EVA operates entirely in latent feature space without relying on explicit future video generation. With approximately 1.2 billion parameters, it is lightweight enough for edge deployment on the robot's onboard computer without cloud dependency.
The model achieved 98.6% success rate on the LIBERO benchmark and improved RoboTwin from 88.9% to 90.3%. Four real-world demos were shown including Klotski puzzle solving, assembly line sorting, and autonomous cleaning.
Commercial Traction: Founded less than a year ago, Xingyuanzhi has raised over 1 billion RMB in funding. It positions as an embodied brain provider rather than robot hardware manufacturer.
