Qiming Venture Partners' portfolio company UBTECH (SEHK:9880) has successfully conducted the world's first collaborative practical training program for humanoid robots at ZEEKR's 5G Intelligent Factory, where multiple humanoid robots seamlessly collaborated across multi-task, multi-scenario industrial environments. This initiative represents a key advancement in developing a general-purpose Swarm Intelligence system for humanoid robots, marking a major leap from single-agent autonomy to Swarm Intelligence. 

Swarm Intelligence
The Next Frontier in Humanoid Robotics

As industrial automation progresses, production-line tasks present new challenges for humanoid robots, particularly in multi-humanoid robot collaboration. Advancing Swarm Intelligence is a critical step in enabling humanoid robots to operate effectively at scale.

UBTECH's Walker S1 humanoid robots are pioneering this shift by deploying a networked cluster to explore how advanced embodied intelligence emerges through swarm behaviors and high-frequency interaction with the physical environment. To support this, UBTECH has developed BrainNet, a software framework for humanoid robot collaboration, and introduced the Internet of Humanoids (IoH), a centralized control hub serving as a blueprint for the software and hardware implementation of Swarm Intelligence.

UBTECH humanoid robots demonstrate "Swarm Intelligence" in collaborative practical training

For humanoid robots to collaborate effectively, UBTECH's BrainNet framework links cloud-device collaborative inference nodes and skill nodes, forming a super brain and an intelligent sub-brain within the Swarm Intelligence system. The super brain, powered by a large reasoning multimodal model, enables intelligent hybrid decision-making and manages complex production-line tasks. The intelligent sub-brain, based on the Transformer model, integrates cross-field fusion perception and multi-robot collaborative control to support parallel distributed learning, accelerating skill generation and transfer.

These innovations allow humanoid robots to move beyond individual task execution, enabling them to collaborate flexibly at the production-line level. This ensures efficient teamwork in complex industrial environments and paves the way for the next evolution of intelligent manufacturing.

To address the high-dimensional decision-making demands of complex production-line tasks, UBTECH has developed the world's first large reasoning multimodal model for humanoid robots. As the core engine of the super brain, this AI system enables the continuous self-evolution of BrainNet, unlocking the full potential of Swarm Intelligence. Currently under development using DeepSeek-R1 deep reasoning technology, the model is being engineered to process data at scale and equip humanoid robots with human-like common-sense reasoning. It allows them to break down, schedule, and coordinate tasks autonomously, optimizing multi-robot collaboration in complex industrial workflows.

UBTECH's multimodal reasoning model is trained on a high-quality industrial dataset accumulated from frontline practical training with the Walker S series across multiple automotive factories. By integrating multimodal features and leveraging Retrieval-Augmented Generation (RAG) technology, the model adapts rapidly to specialized job functions, significantly improving decision-making accuracy, generalization across various workstations, and scalability for large-scale industrial deployment.

Advancing In-Factory Training
Multi-Robot, Multi-Scenario, Multi-Task Collaboration

UBTECH's humanoid robots have now entered Practical Training 2.0, marking their evolution from single-agent autonomy to Swarm Intelligence. At the ZEEKR 5G Intelligent Factory, dozens of Walker S1 humanoid robots have been deployed across complex production areas, including the final assembly workshop, SPS instrumentation zone, quality inspection area, and vehicle assembly station. Working in unison, these humanoid robots successfully executed collaborative sorting, collaborative handling, and precision assembly, demonstrating seamless multi-robot collaboration in real-world industrial settings.

In the collaborative sorting phase, UBTECH's Walker S1 humanoid robots utilize cross-field pure vision-based perception technology and intelligent hybrid decision-making to optimize sorting tasks. Using pure vision-based cross-field perception, the robots continuously track dynamic targets across environments, enabling swarm collaboration through collective mapping and shared intelligence. The intelligent hybrid decision-making system, powered by a large reasoning multimodal model, integrates semantic VSLAM navigation and dexterous manipulation capabilities. This allows for dynamic task allocation between the cloud-based and on-device brain nodes, supporting a cloud-device collaborative decision-making process for Swarm Intelligence.

Multiple UBTECH humanoid robots, Walker S1, collaborate to handle large, heavy-duty containers

In collaborative handling, humanoid robots may face significant challenges, including uneven load distribution, complex trajectory planning, and dynamic environmental adaptation. To address these, UBTECH has developed a joint planning and control system, enabling multi-robot collaboration in trajectory planning, load identification, and compliant control. This ensures robots dynamically adjust their posture and force during handling, significantly enhancing stability and efficiency when transporting large, heavy workpieces.

In precision assembly process, Walker S1 robots demonstrate exceptional dexterous manipulation capabilities, particularly in handling deformable objects. When working with small and deformable film materials, Walker S1 utilizes high-precision sensing and adaptive control technologies to dynamically adjust its grasping force and posture. This ensures film objects remain undamaged and properly aligned throughout the assembly process. This innovation highlights the flexibility and reliability of UBTECH's humanoid robot, which is equipped with dexterous robotic hands with tactile sensing for complex industrial applications.

Additionally, Walker S1 integrates vision-based global initial positioning with force-based secondary positioning via reinforcement learning to execute precision operation-based quality inspection tasks with unmatched accuracy and adaptability.

UBTECH's swarm-intelligent humanoid robots conduct practical training at ZEEKR's 5G Intelligent Factory

UBTECH collaborates with leading industry players, including Dongfeng Liuzhou Motor, Geely Auto, FAW-Volkswagen Qingdao, Audi FAW, BYD, BAIC New Energy, Foxconn, and SF Express. The company's Walker S series humanoid robots are now deployed in more automotive factories worldwide than any other humanoid robot, having successfully completed the first phase of single-agent autonomous intelligence training. With the launch of Practical Training 2.0, UBTECH is accelerating its multi-humanoid robot collaborative training and AI model development to drive product and technology innovation. Beyond Geely Auto, UBTECH is expanding deployment to additional partner factories, further reinforcing the scalable adoption of humanoid robots in industrial applications.