As the post-holiday work surge gains momentum, staff at the Guangdong Embodied Intelligence Training Base in Guangzhou's Haizhu District are methodically calibrating equipment while humanoid robots are being deployed across the facility. A cohort of robotic "trainees" is entering this specialized "school" to undergo pre-employment training for roles in retail, industrial handling, and facility maintenance. As a public service platform supporting Guangdong's embodied AI industry, the training base has nearly completed hardware installation for its integrated control center. Five core platforms—covering operations management, data collection, simulation, data management, and model training—have been successfully interconnected, with full operational capacity across 5,400 square meters of core space expected within the year.
Recent provincial high-quality development meetings emphasized removing barriers for technological innovation and market exploration, aiming to accelerate progress in low-altitude economy, autonomous driving, and practical embodied AI applications. Guangsha Holdings Chairman Lyu Yongzhong highlighted the need to breakthrough in core technologies like high-end components while advancing the provincial embodied AI training base’s development.
The training base represents a critical step toward practical application of embodied AI. What exactly do humanoid robots learn here? How will this "empowerment hub" function? The goal is systematic skill acquisition: moving from lab prototypes to industrial applications requires extensive engineering validation and pilot support. Reflecting this, embodied AI firms have emerged nationwide, with multiple training bases now operational.
Guangdong’s training base leverages local industrial strengths, with its first phase featuring scenarios such as automotive assembly SPS sorting and supply, building and workshop inspections, multi-category parts grasping, commercial retail, and maintenance of power, communication, and computing facilities. Technicians explained that the base’s design accounts for varying operational demands across sectors, incorporating meticulous technical support and setup details.
The project will eventually comprise seven technical platforms functioning as an intelligent assembly line, enabling end-to-end support from data collection and simulation to training, evaluation, and application—systematically teaching robots to work efficiently and improve over time. For instance, automotive SPS sorting, repeated thousands of times daily in real factories, demands extreme precision. In a simulated assembly line at the base, operators remotely train humanoid robots to pick and place components like bolts and fasteners from material racks.
Behind seemingly simple tasks, multiple platforms collaborate: a data collection platform records robot movements; a data management platform cleans, categorizes, and labels this data before transferring it for model training; trained models are then validated on robots. An evaluation platform ensures model quality and reliability, while an integrated operations platform acts as the "brain," coordinating all processes.
Technicians note that manual SPS sorting is prone to fatigue-induced errors and safety risks during night shifts. Using robots for flexible sorting not only enhances safety and efficiency but also enables data-driven optimization of material management and workflows, making processes traceable, quantifiable, and standardized.
At industrial grasping stations, the focus is on "tactile flexibility." Robots equipped with high-density tactile sensors learn varied strategies for handling objects of different materials and shapes—gripping bolts firmly, handling medicine bottles gently, or adjusting holds to prevent slippery tubes from falling. This mimics teaching a child to touch objects, with each tactile experience stored as data, gradually building "muscle memory" to meet flexible, small-batch manufacturing needs.
In patrol scenarios, robots practice generating work orders, opening doors, and disposing of trash. In retail settings, they learn restocking, sorting, packing, and inspecting shelf wear. Collected data will form a high-quality database to empower model training across applications. The base aims to build a "training skill library," offering companies cost-effective algorithms, platform resources, and testing environments while sharing reproducible expertise for typical scenarios.
While infrastructure develops, ecosystem building—a more complex and vital task—is underway. Embodied AI, as a multidisciplinary field, thrives on cross-sector collaboration. Guangdong is adopting a unified approach, creating a "1+1+N" training base system: a main hub coordinating with subsidiary bases under a central management framework. This avoids redundant construction, pools provincial training resources and data, accelerates high-quality dataset scaling, and advances general AI through data-driven, heterogeneous training.
In August 2025, Guangsha Holdings established Guangdong Embodied AI Technology Co., Ltd., which operates the training base and serves as the management center. From the outset, the company positioned itself not merely as a physical space but as an "empowerment hub" and "super-connector" for the industry. Through open capital increases, it attracted nearly ten strategic investors, raising over RMB 100 million to support industrial and technological innovation—a pioneering move for state-owned enterprise synchronization in formation and expansion.
Within six months, the preparation team conducted nearly 300 research sessions across 13 cities, engaging almost 200 ecosystem partners and collaborating with over 20 authoritative institutions for expert reviews and internal seminars. These efforts fostered a collaborative ecosystem spanning infrastructure, robot hardware, and industry applications.
The base also plans to certify subsidiary training sites, with batches of candidates announced during the Spring Festival. This year, Phase 2 expansion will add scenarios in commercial services, industrial manufacturing, intelligent inspection (including specialized operations), home-based wellness, and medical care—covering over 20 sub-sectors such as coffee vending, mine inspections, electronic device testing, home services, and support for adolescents with depression.
The company’s lead stated that the base will gather stakeholders from across the industry chain—some bringing challenges, others product ideas—to help them find ideal partners, streamline collaboration, and accelerate real-world adoption of embodied AI in living and working environments.
Given current industry capabilities, the base will prioritize real-world needs, forming 30–50 innovation alliances across smart manufacturing, healthcare, home-based wellness, inspection, emergency response, and commerce. It will build "industrial project" and "innovation case" libraries, integrating technical empowerment, industry incubation, and supply chain coordination to deepen synergies between advanced technology and traditional sectors.
Comments