Guangxiang Technology has announced the completion of a cumulative seed funding round totaling several hundred million yuan. The latest tranche of this financing saw deep participation from leading financial and industrial investors, including Zhuhai Science and Technology Industry Group, Xingsheng Capital, Songhe Capital, Shunxi Fund, Muhua Science and Technology Innovation, See Fund, Yichen Capital, and the listed company Xingyun Technology. Existing investors such as 01VC and L2F Light Source Entrepreneurs Fund continued to provide additional funding.
The funds raised will be primarily allocated towards the research, development, and iterative enhancement of the company's physically-native foundation model, and to advance the commercial delivery of its embodied intelligent robot products.
Guangxiang Technology is an embodied intelligence company jointly incubated by Tsinghua University's School of Vehicle and Mobility and School of Artificial Intelligence. It is dedicated to constructing self-evolving embodied intelligent agents driven by physically-native intelligence, aiming to promote the industrial application of general embodied intelligence technology. Currently, the company has completed real-world scenario validation for typical high-value workstations in automotive manufacturing, such as loading/unloading and quality inspection. It has also established commercial partnerships with several leading domestic and international automotive enterprises, becoming a pioneer in implementing embodied intelligence within the automotive sector.
Presently, mainstream technological approaches face fundamental limitations in achieving general-purpose physical interaction for robots. The Visual-Language-Action (VLA) approach grafts action experts onto visual language models for task reasoning and action generation. However, semantic generality does not equate to physical interaction capability; such models are essentially mappers of perception to action and struggle to develop general operational abilities through physical interaction. Video-predictive world models focus on pixel-level prediction of environmental observation sequences, but predicting appearances is not the same as understanding physical causality. Images are insufficient to characterize physical properties such as mass, inertia, friction, deformation, and contact, making it difficult to support the generation of generalizable actions.
According to Zhang Tao, Founder and CEO of Guangxiang Technology, true physically-native intelligence is an autonomously emergent capability developed through perception, interaction, feedback, exploration, and constraint within the physical world. Therefore, a physically-native foundation model must prioritize physical interaction as its core principle. It must be capable of continuously learning world laws, behavioral consequences, and task constraints from the physical environment, possessing the general intelligence necessary for complex task completion.
It is reported that the physically-native foundation model independently developed by Guangxiang Technology relies on self-built, high-fidelity, large-scale, interactive physical data assets and a proprietary matrix of cutting-edge reinforcement learning algorithms. Through extensive model interaction within the physical environment, it learns explicit physical law deductions concerning dynamics, contact, constraints, and conservation, as well as implicit environmental state reasoning and physical behavior attribution related to randomness, uncertainty, and long-term consequences. This process enables the model to develop a common-sense understanding of physical laws through continuous physical interaction, thereby achieving truly general and adaptable behavioral capabilities.