Manifold space secures one billion in funding, launches a real-time world model robot with one-tenth of the parameters

According to Beating Monitoring, the robot large-model company Manifold AI, Manifold Space has completed a new round of financing amounting to several hundreds of millions of yuan, pushing the total funding raised in the Pre-A stage to nearly 1 billion yuan. Investors include Guoxin Fund, Yifeng Capital under Temasek, BAIC Industry Investment, and Xineng Venture Capital. Manifold Space was founded by former SenseTime executive Dr. Wu Wei and the Tsinghua University FIB Laboratory, and its core team has won the Waymo Autonomous Driving Simulation Challenge twice.

Its self-developed real-time world model WorldScape serves as the pre-training foundation model for robots, supporting both movement and manipulation dual-interaction. WorldScape adopts a mixture of experts (MoE) architecture. On the evaluation leaderboard WorldScore, it has maintained the top position for nearly two months with only 10% of the parameters of mainstream competitors. By combining the spatiotemporal prediction state of WorldScape, the action model WorldScape Policy can filter out random disturbances such as complex lighting and backgrounds, enabling precise zero-shot control. It has already been deployed in embodied scenarios such as e-commerce logistics and 3C manufacturing.

To test the decision-making capability of large models in real embodied tasks, Manifold Space, together with dozens of research institutions, has launched the evaluation benchmark WorldArena. The evaluation includes 6 major dimensions and 16 indicators. The evaluation focus has shifted from the realism of video frames to the practical value of robot decision-making, and it has already been used to support the CVPR 2026 World Simulator Challenge.