Embodied AI Startup Noematrix Raises Millions, Unveils Force-Centric Brain System
SHANGHAI, CHINA – September 6,2024 – Noematrix, a Chinese startup specializing in embodied artificial intelligence (AI), has announced the completion of a multi-million dollarseed and Pre-A round of funding. The company, founded in November 2023 as a strategic spin-off from robotics company, Flexiv, aims to develop a comprehensive embodied AI brain system, along with related tools and platforms.
The funding round was led by prominent investment funds Prosperity7 Ventures and GF Xinde, with participation from renowned investors such as ZeYu Capital,Innovation Works, Qijic Chuangtan, PlugandPlay China, MFund, and others. The seed round was led by Xiao Miao Lang Cheng, with participation from MFund, Jinghai Zhan Yue, and SEE Fund.
Noematrix’s funding will be used to accelerate product development, expand commercialization efforts, and attract top talent.
The company’s approach to embodied AI focuses on addressing the key challenges of effectively describing the physical world and its interactions with robots. Existing natural language and vision-language models struggle to fully representthe complexities of the real world, relying solely on visual and linguistic modalities. Additionally, training embodied AI models requires expensive and inefficient data acquisition processes, lacking standardized methods.
To overcome these hurdles, Noematrix has developed a unique technological roadmap and strategic approach, centered around a force-centric embodied brain system.This system prioritizes the acquisition and utilization of force data, enabling the development of efficient and cost-effective solutions through rapid iteration of general-purpose skills.
Traditional embodied AI architectures rely on visual-language models for environmental and task representation learning and inference, employing imitation or reinforcement learning to optimize trajectory-based decision-making. Noematrix, leveraging its expertise in force control technology, introduces a multi-dimensional force modality and control strategies into the learning space. This creates a framework for embodied AI models grounded in physical common sense and behavioral decision-making.
Noematrix’s core technology includes:
- Entity World Large Model: This model leverages operation-oriented topological structure common sense understanding, operation-task-centric physical common sense understanding, and a pose-shape joint training framework. It boasts high accuracy, low training costs, and effectively addresses the challenges of flexible object manipulation.
- Robot Behavior Large Model: Recognizing the crucial roleof force perception in robot-environment interactions, Noematrix has developed a force-position hybrid robot behavior large model. This model enables comprehensive force-position hybrid decision control, offering advantages like complete training information, high transferability, precision, and robustness.
The combination of these two models, with the EntityWorld model focusing on task-centric abstraction of operational physical common sense and the Robot Behavior model centered on force and objects, significantly reduces data requirements and enhances the feasibility of real-world applications.
Noematrix’s technological advancements are deeply rooted in the extensive academic and industry experience of its founding team.
Co-founder Dr. Lu Ce-Wu, a professor at Shanghai Jiao Tong University’s Department of Computer Science and Engineering, is a pioneer in the field of embodied AI. During his post-doctoral research at Stanford University’s Artificial Intelligence Laboratory, he was mentored by Professor Fei-Fei Li and Professor LeoGuibas, a member of the National Academy of Engineering.
Dr. Lu is a recipient of the 2018 MIT Technology Review’s 35 Innovators Under 35 award and the 2023 Science Exploration Award for his contributions to embodied AI. His team has madegroundbreaking advancements in the field, including the development of the AnyGrasp grasping algorithm, which surpasses human capabilities, and the HAKE (Human Action Knowledge Engine).
Dr. Lu and his team have published over 100 papers in top journals and conferences, including Nature, Nature Machine Intelligence, and TPAMI. He has served as a reviewer for Science and as a program chair for conferences like CVPR, ICCV, ECCV, and IROS. Notably, his team’s RH20T dataset is the first large-scale embodied dataset to incorporate force information, and Dr. Lu is theonly Chinese scholar invited to join the Open X Embodiment Dataset international collaborative project.
Co-founder Dr. Wang Shi-Quan, who holds a PhD from Stanford University, is currently a visiting professor at Shanghai Jiao Tong University and a standing committee member of the Shanghai Association for Science and Technology. He is dedicated todeveloping general-purpose intelligent robots and founded Flexiv in 2016, pioneering a path towards human-like general-purpose robots based on extreme force control and hierarchical intelligence.
Flexiv’s adaptive robotic products are widely used in various industries, including industrial manufacturing, food processing, and healthcare services. Theyhave pushed the boundaries of robotic capabilities across nearly a hundred applications.
During the 2024 World Artificial Intelligence Conference, Noematrix showcased its embodied AI brain system’s capabilities in real-world demonstrations across various tasks. These included floor cleaning, desk organization, flexible object manipulation (like folding clothes), andfood processing (like peeling cucumbers).
Noematrix also officially launched its self-developed Noematrix Brain, a comprehensive platform and toolkit designed for versatility and intelligence. It covers the entire lifecycle of intelligent agent applications, from development and testing to deployment and validation.
Noematrix’s integrated hardware and software system, optimized for hardware selection and algorithm/software architecture acceleration, ensures exceptional performance and compatibility with various embodied AI devices. The company has transformed its hardware and software expertise into standardized solutions, offering an application development platform that integrates body hardware, basic software frameworks, algorithm models, and supporting tools. This empowers developers to efficiently and convenientlycreate embodied AI applications.
Through a unified model architecture, in-depth business insights, and careful scene selection, Noematrix has identified the critical need for breakthroughs in common skills to address market pain points. The company has introduced the AnySkill series of general-purpose skills, compatible with various embodied AI platforms. Theseskills have been rigorously tested in real-world scenarios, demonstrating high environmental adaptability, robust object generalization capabilities, and significant market value.
The extensive accumulation and combination of general-purpose skills within the Noematrix Brain significantly enhance its algorithmic capabilities, expanding its application scope.
Noematrix has already demonstrated a gradual pathfrom technological development to commercialization through the application and business exploration of its AnySkill series. This validation not only proves the effectiveness of the technology but also showcases its potential for sustained growth. As technology advances and commercial applications expand, Noematrix’s embodied AI brain is poised to play a significant role in various fields,driving the widespread adoption of embodied AI technology.
Views: 3