One Brain, Multiple Forms: From Concept to Real-World Deployment
From August 19 to 23, MagicLab made its appearance at WRC 2026. Centered on the theme of “Building a Scenario-Driven, Physical AI-Native Platform,” the company brought its “Magic Town” to the exhibition floor, immersively showcasing solutions for real-world business scenarios including industrial manufacturing, logistics sorting, and public safety. Through three major demos, it presented the latest advances of its self-developed general-purpose embodied foundation model, Magic-VLA K02, and broke down the technology loop of “scenarios pose the questions—models solve the problems—data feeds back into the system.”
During the conference, MagicLab also hosted the “Scenario-Driven, Industry-Demand Integration” forum, leveraging the opportunity to bring together leading resources from industry, academia, and research. It worked with upstream and downstream companies across the industrial chain to explore pathways for scaling embodied intelligence—from demonstrations to practical applications, from isolated deployments to ecosystem-wide collaboration, and from regional coordination to global competition. The forum showcased commercialization results and unleashed powerful ecosystem momentum.
Immersive Recreation of Real-World Business Operations: Industrial Manufacturing, Logistics Sorting, and Public Safety Become the Exhibition’s Top Highlights
At this year’s WRC, closed-loop validation in real-world business scenarios became the key benchmark for evaluating the capabilities of embodied-intelligence companies. At the exhibition, MagicLab brought real-world task scenarios from factories, warehouses, and industrial parks into the venue, allowing visitors to see, within a single space, the real engineering applications and commercial value of embodied intelligence in industrial manufacturing, smart logistics, public safety, and other industries.
At the exhibition, MagicLab’s industrial wheeled humanoid robot MagicBot D1 was undoubtedly the center of attention. Combining long-range mobility with fine manipulation, it smoothly completed the full workflow of material handling, loading and unloading, and cross-area transfer beside a simulated cargo hold.
D1 adopts a hybrid configuration combining a wheeled chassis with humanoid upper limbs. It is equipped with a 7-degree-of-freedom force-controlled robotic arm, with a single-arm payload of 5 kg and force-control accuracy of 0.5 N. It also features a centimeter-level mapping and localization system—with a positioning error of no more than 7 cm—and a self-developed RCS fleet-dispatch system. Its operating height covers 80–230 cm, with a reach radius of 90 cm. It supports hot-swappable batteries and automatic recharging, enabling unattended operation 24/7.
Positioned as an industrial-grade intelligent robot designed for mass production, the product focuses on replacing repetitive manual physical labor. D1 has already taken on real production tasks at Dreame’s smart manufacturing plant, including workshop material handling and loading and unloading on production lines, where it has undergone reliable validation. It is also becoming a core vehicle for MagicLab’s future solutions in industrial manufacturing and smart logistics.
In industrial manufacturing, MagicLab has established a multi-form deployment path based on “one brain, multiple forms.” The same general-purpose intelligent brain can drive different hardware forms—including humanoid robots, wheeled humanoid robots, and sorting robots—to take on different roles at scale. This fundamentally reduces the training and development costs associated with transferring capabilities across robot embodiments, enabling embodied intelligence to become truly integrated into core manufacturing processes.
In smart logistics, MagicLab recreated the complete business workflow of warehouse circulation and demonstrated collaborative operation among multiple modular products. The wheeled humanoid robot XGZ1 and the sorting robot HyperBot performed heterogeneous multi-robot collaboration, presenting a complete closed-loop solution covering the entire “package induction—sorting—transfer” process.
At the booth, XGZ1 served as a flexible workstation. Using 3D vision and multimodal AI-based perception, it precisely picked up disordered, randomly stacked materials one by one and placed them into the HyperBot sorting system for continued processing. HyperBot moved at a speed of 2.5 m/s and navigated smoothly through curved turns, accurately diverting goods to their designated locations and prompting crowds of visitors to raise their phones and record the demonstration.
With XGZ1’s six-second end-effector quick-change system, the solution breaks through the rigid “one machine, one purpose” model of traditional industrial robots. Its retrofit-free compatibility reduces infrastructure investment. Powered by MagicLab’s self-developed RCS dispatch system, it enables fleet-level collaboration and data-link integration among humanoid robots, robotic arms, AGVs, and other robot forms. This creates a data flywheel covering “equipment deployment—on-site operation—algorithm iteration,” providing the logistics industry with modular, immediately deployable solutions for its transition from rigid automation to flexible intelligence.
In public safety, MagicLab’s “smart policing” solution has been deployed in core business scenarios including traffic management, patrol and emergency response, and government services.
Previously, MagicLab worked with the Wuxi Municipal Public Security Bureau to deploy a traffic-management robot at the 2026 Wuxi Marathon. The robot successfully completed real-world tasks including traffic direction and crowd control, becoming the world’s first bipedal robot traffic officer to serve at a major sporting event and earning public praise from a spokesperson for China’s Ministry of Foreign Affairs. In July this year, the company took the lead in establishing the Jiangsu Key Laboratory of Embodied-Intelligence Robots for Public Safety. It built an industry–academia–research–application collaborative innovation consortium between leading enterprises and police colleges, earning official recognition for its technical capabilities and practical value.
Drawing on its accumulated capabilities in customized policing R&D and end-to-end delivery, MagicLab is working with ecosystem partners to deeply integrate with China Mobile Lianyungang Branch’s 5G-A computing-power network infrastructure. The parties are conducting joint research focused on frontline policing scenarios such as traffic management, emergency response, and government services.
Public safety is a real-world “proving ground” for embodied-intelligence technology. Leveraging its combined strengths in full-stack self-developed technology, frontline operational experience, and support from a key laboratory, MagicLab has taken the lead in establishing a closed loop for the large-scale application of humanoid robots in public-security operations, accelerating the transition of embodied intelligence from laboratories to real-world business environments.
In the future, MagicLab will continue developing service robots based on practical needs. Beginning with three major commercial scenarios—hotels, laundries, and shopping malls—the company will work with ecosystem partners to build integrated solutions covering the full range of business processes, including reception and guided shopping, item sorting, unmanned operations, and terminal delivery. These solutions will empower the intelligent upgrade of offline services, help physical retail move beyond the traditional static unmanned-cabinet model, and usher in a new era of mobile embodied-intelligence services.
Completing Its Quadruped Robot Product Portfolio for Industrial Inspection: Lightweight Industrial Quadruped Robot MagicDog T1 Makes Its Grand Debut
At this year’s WRC, MagicLab unveiled the lightweight industrial quadruped robot MagicDog T1. The addition further expands the company’s product portfolio and provides a new standardized product module for solutions in scenarios such as industrial inspection and industrial-park security.
T1 is built around the product philosophy of “lightweight yet heavy-duty, with intelligent legs leading the way.” Together with the industrial-grade quadruped robot MagicDog Y1, it creates a light-to-heavy payload spectrum. Compared with Y1, T1 is smaller, more cost-effective, and more agile. Its core strengths lie in modularity and scenario adaptability, making deployment feasible in light-industrial environments such as industrial parks, warehouses, and retail stores—scenarios where industrial quadruped robots were previously considered too expensive or unnecessary.
In terms of performance, T1 weighs approximately 18 kg, supports a maximum continuous payload of 15 kg, and reaches a top speed of 5 m/s. It can climb steps up to 18 cm high and handle 40° slopes with ease. It supports transportation and rapid deployment by a single person. The robot is equipped with an RGB camera, RGB-D vision, and a 16-line LiDAR, providing autonomous navigation, intelligent obstacle avoidance, target tracking, and real-time video transmission. Its maximum continuous operating time without a payload is approximately three hours, and it supports both automatic recharging and battery swapping.
In particular, T1 can quickly accommodate various gripper attachments and industry-specific modules through standardized expansion interfaces on its back, enabling lightweight tasks such as picking up trash, pressing buttons, and clearing debris. Inspection and on-site intervention are integrated into a single platform. A robot dog is no longer merely a “camera that can walk,” but a multipurpose platform capable of carrying out hands-on work. This modular design aligns precisely with the flexible demands of industrial scenarios for “one machine with multiple capabilities and on-demand expansion.”
T1 has already undergone practical validation at Dreame’s smart manufacturing plant. Equipped with a robotic arm, it completed lightweight tasks such as clearing debris from the floor and pressing equipment buttons. At this year’s WRC, T1 demonstrated lightweight operations including floor-debris removal with a robotic arm. Together with vision and environmental sensors installed through its rear expansion interface, it performed integrated tasks including equipment-status inspection, button operation, and floor-debris removal, further expanding the range of tasks quadruped robots can undertake in factories.
Magic-VLA K02 General-Purpose Embodied Foundation Model Evolves Across the Board: “One Brain, Multiple Forms” Moves from Concept to Engineering Reality
How can different hardware forms—including quadruped robots, humanoid robots, and wheeled robots—work together to tackle complex tasks in real-world scenarios? MagicLab’s answer is a general-purpose embodied foundation model that serves as the central “brain” of physical AI.
At WRC, MagicLab presented three demos showcasing the capabilities of its Magic-VLA K02 general-purpose embodied foundation model, attracting visitors who stopped to watch and interact.
In the demonstration area for intangible cultural heritage woodblock printing, the robot coordinated both hands to perform a sequence of actions—applying ink, placing paper, pressing it with a weight, and handing over the finished product—in one seamless flow, producing customized cards for visitors. This demo showcased Magic-VLA K02’s capabilities in long-horizon task planning and execution, validating heterogeneous bimanual collaboration and precise temporal action planning.
With traditional industrial robotic arms, this type of bimanual collaboration requires engineers to manually write extensive coordination logic. Magic-VLA K02, however, uses a high-level system to autonomously decompose the abstract instruction “make a woodblock print” into a sequence of atomic tasks with explicit temporal relationships, reducing the accumulation of errors in multi-step operations at the source.
The demo in the flexible garment-folding area addressed the industry-recognized challenge of manipulating highly deformable objects with many degrees of freedom. After mastering T-shirt folding at WAIC, Magic-VLA K02 faced a previously unseen garment category—a polo shirt—and autonomously recombined existing skill modules in a zero-shot manner. Through multiple rounds of exploration and trial and error, it learned a folding strategy, fully demonstrating the capability leap from “single-item specialization” to “cross-category generalization.”
Visitors could interrupt the folding process at any time. Based on its “key-result image” mechanism, the robot re-identified the garment’s state, assessed deviations from the task, and resumed the subsequent operations.
The box-folding and tape-sealing demo showcased the industry’s first complete execution of a combined long-horizon task involving box folding and sealing by a general-purpose embodied foundation model. It validated the combined challenge of fine manipulation of rigid boxes and dynamic control of flexible tape. With an on-site success rate of over 90%, the demonstration marked the transition of large-model-driven long-horizon manipulation from the laboratory toward deployable productivity tools.
Although the three demos differed significantly in their physical task properties, what they presented were not three independent fixed programs, but the externalized capabilities of the same model framework across different tasks. In April this year, MagicLab officially released its self-developed general-purpose embodied foundation model, Magic-VLA K02. After four months of development, the model has undergone comprehensive evolution. Its hierarchical dual-system collaborative architecture has brought overall accuracy on complex long-horizon tasks to 92%, reduced the task-interruption rate to below 5%, increased scenario-adaptation throughput by 110%, and achieved a 100% success rate for cross-device adaptation. It has also reduced the amount of robot demonstration data required for training by 60%, providing the industry with quantifiable technical benchmarks for moving from “single-item specialization” toward “general-purpose generalization.”
By introducing a metadata description system, Magic-VLA K02 provides a unified representation of different robots’ embodiments, control modes, and action spaces. This allows the same model framework to support robotic arms, mobile manipulation robots, and different hardware configurations such as single-arm and dual-arm systems. Device-category compatibility has increased by more than 200%, moving “one brain, multiple forms” from concept to engineering reality.
Named One of the 2026 Global Top 100 Benchmark Embodied-Intelligence Companies: MagicLab and Its Ecosystem Partners Discuss the Industry’s Next Decade
“Real work, real intelligence, and real-world deployment” became the defining theme of this year’s WRC. The commercial deployment of embodied intelligence was the industry’s most closely watched topic. To date, MagicLab’s orders on hand exceed RMB 1.1 billion, making it a benchmark company in the industry’s steady progress toward mass production, delivery, and engineering deployment.
On August 21, MagicLab was named one of the 2026 Global Top 100 Benchmark Embodied-Intelligence Companies. MagicLab CTO Chen Chunyu was invited to attend the Embodied-Intelligence Commercialization Forum hosted by EO Intelligence, where he discussed mass production, delivery, and engineering deployment of embodied-intelligence robots with leading companies in the industry.
Chen Chunyu believes that mass production is a systematic test. Scaling from hundreds of units to thousands tests the supply chain, engineering capabilities, and stability verification. The ultimate goal is to reach a scale of more than 10,000 units, enabling robots to serve customers reliably and stably over the long term. Looking ahead, a general-purpose brain is the ultimate goal. Before that arrives, companies still need to develop deeply in a large number of constrained scenarios. MagicLab will remain focused on full-stack self-development, practice a scenario-driven approach, and work with partners from all sectors to explore the industry’s development together.
In July this year, MagicLab was unanimously elected as the second rotating chairing institution of the Jiangsu Embodied-Intelligence Robot Industry Alliance, succeeding the inaugural chairing institution, Estun Automation. The handover was viewed by the industry as a landmark signal of Jiangsu’s robotics industry transition from industrial automation to AI-driven embodied intelligence, and as a reflection of regional industrial resonance in the “first year of physical AI.”
After taking over the alliance, MagicLab released its work plan, “Bridging the Past and the Future, Breaking Through and Setting Standards.” The plan focuses on five areas: scenario deployment, data closed loops, standards development, overseas product expansion, and industry–academia integration. It responds to the supply-and-demand matching mechanism of “local governments solicit needs, the alliance posts challenges, and companies undertake them,” promoting the transformation of embodied-intelligence products from “exhibits” into “commercial products.”
“Going it alone is not enough to build momentum; ecosystem collaboration is the key to large-scale deployment,” said MagicLab President Gu Shitao. “MagicLab is willing to use full-stack technology and a scenario-driven approach to bridge the final mile of deployment, moving forward alongside more than 300 alliance members.”
During WRC 2026, the 2026 Embodied-Intelligence Robot Industry Ecosystem Forum, hosted by MagicLab under the theme “Scenario-Driven, Industry-Demand Integration,” was successfully held in Beijing. Experts and scholars from industry, academia, and research institutions—including the Chinese Academy of Sciences, Tsinghua University, the China Academy of Information and Communications Technology, and the China Center for Information Industry Development—gathered with representatives of upstream and downstream companies in Jiangsu’s embodied-intelligence industry chain. They engaged in in-depth discussions on cutting-edge industry trends, scenario deployment, standards development, and internationalization pathways for embodied intelligence.
“Consumers do not demand general-purpose capabilities, but they do require robots to achieve a task success rate of 99%,” emphasized Huo Jianghao, MagicLab’s head of technology, during a roundtable discussion. “The greatest challenge facing embodied intelligence is not the technology itself, but the ability to deeply integrate deployment, execution, and real-world scenarios.”
The forum was guided by the Jiangsu Provincial Department of Industry and Information Technology and organized by the Jiangsu Embodied-Intelligence Robot Industry Alliance. During the forum, the alliance signed strategic cooperation agreements with TÜV SÜD and TÜV Rheinland, two internationally authoritative certification organizations. This marked a key step forward for Jiangsu’s embodied-intelligence industry on the path toward standardization and internationalization.
At the forum, MagicLab President Gu Shitao said that the company is positioned as a scenario-driven, physical AI-native platform. It remains committed to full-stack self-development of software and hardware, with more than 90% autonomy and controllability across its core algorithms and hardware. The company has achieved commercial deployment in industrial manufacturing, smart logistics, public safety, and other scenarios. Its products are available in nearly 30 countries, with overseas revenue accounting for 40%. The first year of physical AI has arrived. Going forward, MagicLab will replace isolated competition with an ecosystem mindset, make scenarios the proving ground and collaboration the accelerator, and work with industry partners to define the next decade of embodied intelligence.