RoboScience has closed a pre-A round of financing worth several hundred million yuan, with its VLOA—integrating embodied world models and large-scale embodied manipulation models—driving the scalable deployment of embodied intelligence.

Release Date:

2026-02-12 00:10

Source:

RoboScience has completed a pre‑A round of financing totaling several hundred million yuan. The round was led by Puhua Capital, with participation from Dacheng Capital, Changshi Capital, Hong Kong WiseTech Innovation, and Tianqi Capital. Existing investors, including China Merchants Capital and 01 Venture Capital, also increased their stakes. Huaxing Capital served as the exclusive financial advisor. This funding will be primarily used to further refine the company’s core VLOA large‑model technology, thereby accelerating the realization of its vision to build general‑purpose robots. (The VLOA integrates embodied world models and large‑scale embodied manipulation models, driving the scalable deployment of embodied intelligence.)

RoboScience has launched an end‑to‑end VLOA (Vision‑Language‑Object‑Action) large model that integrates an Embodied World Model and an Embodied Manipulation Large Model, aiming to create a general‑purpose intelligent system applicable to any task, any object, and any robot.

1. Embodied World Model

As a world model decoupled from robotic hardware platforms and rigorously adhering to the laws of physics, its core is understanding the physical world and generating executable future plans.

Data foundation: The model is pre-trained on massive internet video data. Through a fully automated data annotation and cleaning pipeline, we have accumulated over 1 million hours of object‑centric, high‑dimensional multimodal manipulation data (tens of millions of video clips), growing at a rate of hundreds of thousands of hours per week. Our goal is to build a world‑leading dataset of tens of millions of hours by 2026.

Core Competency: a. Ensure that all mission planning complies with real-world physical constraints;

b. Forecasting future sequential tasks by modeling the probability distribution of multiple solutions and their corresponding prediction confidence levels;

c. Achieved temporal and spatial global consistency in trajectory prediction for complex long-term tasks.

Enabling robots to autonomously perform complex, fine-grained tasks requires more than just an embodied world model; robots cannot truly learn physical laws from video alone. Therefore, we have also developed a large-scale embodied manipulation model, leveraging vast amounts of data generated by our in-house multimodal physics simulation engine to learn physical principles.

2. Embodied Large-Scale Models for Manipulation

This model addresses the challenges of generalization and dexterous manipulation through an efficient closed-loop workflow that integrates a physics engine, simulation data, and end-to-end training.

Data foundation: Using a self‑developed multimodal physics engine, we have accumulated a 10B high‑quality manipulation dataset (10 billion manipulation actions on objects in full 3D space). Our 2026 goal is to build a 1T high‑quality manipulation dataset (1 trillion manipulation actions).

Core Capabilities: a. Supports a wide range of manipulation tasks for objects across all spatial dimensions, including rigid bodies, hinge‑connected bodies, and 1D/2D/3D deformable bodies;

b. Cross-ontology, supporting various types of robots and end-effectors;

c. Supports closed-loop operation;

d. Supports multimodal perception data, including visual, tactile, and force‑sensing modalities.

Data Scale Comparison

Ye Tian , founder and CEO of RoboScience, earned his bachelor’s degree from the School of Physics at the University of Science and Technology of China and his master’s degree from Stanford University’s AI Lab, where he studied under Andrew Ng. He previously served as the Technical Lead for Apple’s AI Platform. At Apple, he spearheaded the development and deployment of several groundbreaking AI technologies, including a core infrastructure platform hailed as “Apple’s PyTorch and CUDA,” Apple Intelligence—the world’s first edge‑cloud collaborative large‑model inference system—along with the first on‑device inference system, a multi‑compute‑unit collaborative computing framework, the first compiler‑fusion system, and the first on‑device dynamic neural network. These innovations helped build Apple’s cross‑platform AI ecosystem, enabling the launch of thousands of apps that serve over one billion users across two billion devices, while amassing extensive experience in large‑scale edge AI deployment and ecosystem construction.

Lin Shao, Founder and Chief Scientist, is an Assistant Professor in the Department of Computer Science at the National University of Singapore. He earned his Ph.D. from the Stanford AI Lab, where he studied under Jeannette Bohg and was co‑supervised by Leonidas J. Guibas. The deep neural network architecture UniGrasp, proposed by Lin Shao, has become a benchmark approach for data‑driven dexterous hand grasping. His cross‑embodiment dexterous grasping method D(R,O)Grasp was awarded the Best Paper on Robot Manipulation and Locomotion at ICRA 2025—the first time in five years that an Asian institution has won this honor as the primary affiliation. As early as 2020, he initiated and completed the Concept2Robot project, exploring the integration of natural language and video for learning robot manipulation tasks, one of the pioneering efforts in the VLA domain. Additionally, his SAM‑RL method was nominated for the RSS 2023 Best Systems Paper Award.

Penghai Liu , co‑founder, previously served as Vice President at Ecovacs Group, bringing more than 20 years of experience in new product development and launch. He built Ecovacs’ product development process and integrated supply‑chain management system from the ground up, managed a team of over 3,000 employees, and oversaw the mass production of more than 50 robotic products. Tao Wang , co‑founder, holds a degree in Statistics from the University of Science and Technology of China and formerly led fundraising at SenseTime’s Guoxiang Capital, spearheading the mobilization and deployment of multi‑billion‑yuan industry funds. He also brings over a decade of experience in private equity and investment banking. The company’s core team comprises seasoned professionals from leading global manufacturers, top AI and robotics firms, and world‑class research institutions.

RoboScience is dedicated to building world‑leading embodied intelligence large models and foundational products. In the future, our embodied large models will enable versatile robotic platforms tailored to diverse applications, widely deployed across retail, logistics, industry, and home settings, delivering safe and intelligent solutions.

Embodied World Model — Introducing the VLOA Large Model (Part 1)

RoboScience has released RoboMirage, enabling high-precision unified simulation of contact forces among rigid bodies, soft bodies, and articulated systems.

Embodied World Model — Introducing the VLOA Large Model (Part 1)

RoboScience has released RoboMirage, enabling high-precision unified simulation of contact forces among rigid bodies, soft bodies, and articulated systems.

北京机科未来科技有限公司

RoboScience has closed a pre-A round of financing worth several hundred million yuan, with its VLOA—integrating embodied world models and large-scale embodied manipulation models—driving the scalable deployment of embodied intelligence.