Robot AI startup Manifold AI (流形空间) has raised another round of funding worth hundreds of millions of yuan, pushing its total Pre-A funding to nearly 1 billion yuan ($140 million). Investors include China Reform Fund, Temasek's Yifeng Capital, BAIC Industrial Investment, and Xinneng Venture Capital. The company was founded by Dr. Wu Wei, a former executive at SenseTime, in collaboration with Tsinghua University's FIB Lab. The core team has won the Waymo Autonomous Driving Simulation Challenge twice.
The startup's self-developed real-time world model, WorldScape, serves as a pre-trained foundation model for robots, supporting both mobile and manipulation tasks. WorldScape uses a mixture-of-experts (MoE) architecture and has topped the WorldScore benchmark for nearly two months with just 10% of the parameters of competing systems. By combining WorldScape's spatiotemporal prediction states with its action model, WorldScape Policy can filter out random disturbances like complex lighting and background noise, enabling zero-shot fine manipulation. The system is already deployed in embodied scenarios such as e-commerce logistics and 3C manufacturing (computer, communication, and consumer electronics).
To evaluate large models' decision-making in real-world embodied tasks, Manifold AI has partnered with dozens of research institutions to launch a benchmark called WorldArena. The evaluation covers 6 dimensions and 16 metrics, shifting the focus from visual realism to practical decision value. WorldArena is already being used to support the CVPR 2026 World Simulator Challenge.