Robotics Startup "Rongmian Kaiwu" Secures Multi-Million Yuan Funding for World Model Development
Rongmian Kaiwu (Beijing Rongmian Robot Co., Ltd.), a company focused on embodied artificial intelligence and world models, has successfully completed two seed funding rounds, raising several hundred million RMB. Investors include Dingfeng Kechuang, Yuantu Future, Baidu Ventures, Woyan Capital, Wuyuefeng Kechuang, and Wanlin International, with a new round currently in negotiation. The initial funding was allocated to the research and development of their proprietary world model, LaMPA, the construction of a reinforcement learning system, and the enhancement of data loops and product delivery capabilities.
Founded in March 2026, Rongmian Kaiwu aims to develop foundational models that understand the physical world, predict environmental changes, and enable robots to perform tasks, moving them from single-scenario capabilities towards cross-scenario generalization. The founding team comprises former members from Tsinghua University's autonomous driving background, with significant experience in developing and deploying industry-first intelligent driving world models and embodied reinforcement learning. Key figures include Dr. Xiao Zhongyang, who led the delivery of the industry's first complex interaction scenario world model used in over 700,000 Nio vehicles; Dr. Zhong Yuanxin, a former "Huawei Genius Youth" who designed and mass-produced Huawei's next-generation intelligent driving world model; and Dr. Wang Yunlong, who worked on large model and real-robot reinforcement learning at Ximone Robotics.
The company's world model, LaMPA, addresses the need for robots to understand physical world dynamics beyond imitation learning. LaMPA employs a three-tiered representation system: environment, ego (robot's own state), and experience (prior knowledge of object properties and affordances). This forms a latent space for the base model to learn causal relationships, predict future states, and generate robot control actions. LaMPA also utilizes a Block Diffusion structure for efficient training and scalability. To overcome limitations in traditional reinforcement learning, Rongmian Kaiwu developed a generalized World Reward Model, distilled from the base world model, to provide stable feedback and accelerate deployment in new scenarios. The company has a strategic partnership with Yuantu Future to implement its solutions in high-precision industrial assembly for server manufacturing, with plans to expand across various production stages and lines. Their data strategy involves self-collection, crowdsourcing, and model-augmented data generation, with a focus on closing the data loop for continuous model iteration and long-term scaling across industrial applications.
This funding round underscores the significant investment flowing into embodied AI and foundational world models for robotics. The emphasis on a "world model" approach, inspired by theories like JEPA, signals a shift from task-specific AI to systems capable of more general understanding and adaptation in the physical world. The challenge lies in translating theoretical frameworks into robust, scalable, and cost-effective solutions for complex industrial environments. The company's strategy of building a generalized reward model and a standardized product offering, rather than just a model, addresses the critical need for rapid deployment and continuous learning in manufacturing. Future success will depend on the company's ability to demonstrate tangible improvements in robotic generalization, operational efficiency, and cost reduction compared to existing automation solutions, navigating the inherent complexities of real-world physical interaction and data acquisition.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.