A Fudan professor teams up with a former Intel chief scientist to start a business, secures additional angel-round investment, and has partnered with Xiaomi and Unitree Robotics | Exclusive from 36Kr

Author: Ou Xue

Editor: Yuan Silai

Yingke learned that recently, Shanghai Moushen Intelligence Technology Co., Ltd. (hereinafter referred to as "Moushen Intelligence"), the world's first company focusing on generative general embodied brain, completed an angel - round additional investment of over ten million RMB. The investor is Xuhui Capital.

The funds from this round of financing will be mainly used for computing power procurement, expansion of the engineering team, and construction of the robot experiment platform. Meanwhile, the company officially moved into "Mosu Space", the core carrier of the artificial - intelligence industry in Xuhui District.

Founded in 2025, Moushen Intelligence focuses on the research and development of generative cross - ontology general embodied brains, aiming to provide generalizable and implementable "brain - level" capabilities for humanoid robots and multi - form physical agents.

Yingke learned that the Moushen Intelligence team combines the research capabilities of academic institutions and the industrialization experience of global technology companies. It was co - founded by Chen Tao, a professor at the School of Future Information and Innovation of Fudan University and the director of the Deep Learning Laboratory, Zhang Yimin, the former chief scientist of Intel China, and serial entrepreneurs. Among them, Chen Tao once worked on large - model compression and chip adaptation at Huawei Hisilicon in Singapore.

Chen Tao, the founder and chief scientist of Moushen Intelligence, said that the technical route of Moushen Intelligence started in 2021, earlier than the popularization of relevant concepts in the industry. Its first - generation model has achieved the ability to generate spatial actions.

The team's first - developed "action primitive" technology breaks down complex action sequences into thousands of basic elements. The model can dynamically call and combine these primitives according to natural - language instructions to generate new, untrained action sequences, achieving strong generalization ability.

In terms of data usage and training efficiency, the team adopted a unique three - stage training architecture: pre - training the model with about 90% of open - source Internet human - motion videos to enable it to understand general motion laws; then fine - tuning and adapting it with about 10% of simulator data; finally, calibrating it through reinforcement learning with only about 5% of real - machine data. This method significantly reduces the dependence on expensive real - machine data and shortens the adaptation cycle of new ontologies to a few weeks.

The team also has profound experience in model lightweighting and edge - side deployment. Through self - developed model compression and inference acceleration technologies, the number of model parameters can be compressed to 1/4 of the original, and the inference speed can be increased by 10 times, enabling the efficient operation of models with tens of billions of parameters on robot edge - side chips. Relevant research results have been published and won awards at top - level conferences such as IJCAI.

The three - dimensional perception and interaction capabilities are provided by the LL3DA model. This model directly processes three - dimensional point - cloud input and supports multi - modal instructions such as language, images, and videos, enabling robots to immerse themselves in understanding the environment and complete task planning.

(Image source: Company)

Currently, the company has pioneered the industry's first "World Motion Model", enabling robots to react in real - time according to the external environment, achieving environment - immersive action generation and moving towards general physical intelligence. It is reported that Moushen Intelligence is the only team in the industry that has achieved full - link implementation of embodied intelligence, including action generation, three - dimensional world model, model compression, and hardware - software collaboration.

(Image source: Company)

Moushen Intelligence's technology has entered the industrial verification stage and has reached business cooperation with many leading industrial customers such as Unitree Technology, Guodi Center, Hechuan Technology, and Xiaomi Group. In terms of performance, the company has obtained order revenues of over ten million RMB. The current orders mainly come from its ability as a "data generator", providing action - data generation services for customers.

In terms of implementation scenarios, the company clearly focuses on two vertical fields: industrial logistics handling and home health care for the elderly. In the handling scenario, the model needs to meet the requirements of dynamic obstacle avoidance and multi - robot collaboration; in the health - care scenario, it plans to provide progressive services from safety monitoring to limb assistance.

Regarding the business model, Moushen Intelligence aims to become a general "brain" supplier in the robot field. The company hopes to provide an efficiently deployable intelligent core for diverse robot bodies globally, rather than manufacturing hardware itself.

Chen Tao believes that the industrial ecosystem of embodied intelligence will tend to be based on division of labor and cooperation, and it is difficult for a single company to monopolize the entire stack. Moushen Intelligence aims to focus on the core layer of the "World Motion Model". Through its lightweight and highly generalizable technology, it will empower downstream body manufacturers and integrators to jointly promote the large - scale implementation of physical intelligence.

"We focus on breakthroughs in core technologies of edge - side lightweighting and cross - ontology embodied brains," Chen Tao said. Moushen Intelligence will take the "World Motion Model + Edge - side + Scenarios" as the core path to continuously promote technological breakthroughs and industrial integration.

Investor's view:

Ma Xi, the general manager of Xuhui Science and Technology Innovation Investment and the person - in - charge of the incubator fund, said: "As an embodied - intelligence enterprise focusing on the transformation of scientific and technological achievements of Fudan University with brain algorithms as the core technology, Moushen Intelligence's R & D team has the ability advantages and rich R & D experience in the fields of machine vision and multi - modal embodied models. Xuhui Capital's investment is expected to promote the agglomeration of the regional embodied - intelligence industry, strengthen the industrial chain of embodied intelligence in Xuhui District, and support the transformation of scientific and technological achievements of well - known universities."