At the 28th Beijing International Science and Technology Expo - Future Industries Promotion Conference held in Beijing on May 8, 2026, Mao Shijie, Vice President of Lenovo Group, delivered a speech.
The following is a transcript of his speech:
Mao Shijie: Thank you, everyone. I am Mao Shijie from Lenovo Group. It is a great honor to share with you today Lenovo's latest advancements in the field of robotics. Many people are still unaware that Lenovo has been developing robotics for several years. Last year, we formally established a robotics laboratory dedicated to this field.
The topic of robotics is not new; it has been a long-term discussion. Years ago, we saw robotic arms replacing workers in various operations—what we termed Robotics 1.0. Later, with technological advancements like SLAM, robots gained mobility, exemplified by household cleaning robots and hotel delivery robots, which primarily utilized SLAM combined with CNN-based detection and recognition algorithms.
The recent surge in robotics popularity is fundamentally driven by the capabilities brought by Transformer architecture to AI, empowering robots. This evolution has shifted robots from performing fixed-environment tasks to operating in open environments and handling multiple tasks, significantly expanding their application scope. We all envision humanoid robots and future Robotics 4.0—general-purpose robots that do not require pre-set tasks. However, the foundational technical models for such robots are still under exploration. Many companies are actually working on Robotics 3.0 while promoting it as 4.0, so it's important to clarify this basic concept.
We categorize core robotics technologies into several parts. Currently, the most critical and mature is the MPC model predictive control combined with reinforcement learning control. A notable example is the impressive motion control demonstrated by Unitree at this year's Spring Festival Gala. This area is quite mature. Over the next 2-3 years, the industry's focus will be on VLA models, which are gradually moving toward productization. Meanwhile, the academic focus is on world models. We believe world models differ fundamentally from current architectures, as they involve understanding and predicting the real world, representing a necessary path toward artificial general intelligence. However, this remains in the laboratory stage and has not yet reached industrial application.
Based on these industry trends, Lenovo Group positions itself around scenario-driven pragmatic AI. Today's discussion on scenarios implies four aspects:
First, in terms of architecture, we believe mature products today should adopt VLA-like architectures, integrating vision, language, and reinforcement learning—essentially different levels of models working together.
Second, a core aspect is the shift in paradigm to data-driven models on new architectures. Thus, accumulating relevant data in vertical scenarios to drive these models becomes a crucial task.
Third, deploying robots today is more complex than deploying laptops or computers. Implementing robots in a station involves significant challenges, necessitating the development of agile tools and platforms.
Fourth, both VLA and world models have many areas worth exploring. Combining these four aspects forms Lenovo Group's robotics strategy.
Next, I will briefly introduce our progress in several areas, starting with hardware. I am often asked if robots must be humanoid. From today's perspective, practical products are scenario-based and come in various forms. For instance, we believe the most practical trend this year is transitioning from quadruped to wheeled-legged designs, as wheels—though not evolved in nature—offer high efficiency in practical applications.
This year, we will launch the MC small wheeled-legged robot and the MX all-terrain industrial-grade wheeled-legged product. We are also developing an explosion-proof version, EX, based on MX, for use in oil fields and chemical industries. Additionally, we have created a humanoid robot, LX, but we believe humanoid robots are not yet at a general-purpose stage. Therefore, our humanoid robot applications are primarily focused on B2B use in Lenovo's own stores and factories.
Beyond hardware, the biggest challenge in robotics lies in the "cerebellum and brain," especially the "brain." As a researcher in this field, an important part of my job is explaining to clients and leaders that current robots are still quite limited. It is essential to understand that robots consist of a cerebellum and a brain. Most progress so far has been in the cerebellum, with many issues remaining to be solved for the brain.
This year, the brain needs to address spatial intelligence. Robots must have a deeper understanding of their environment, beyond merely detecting obstacles with radar or point clouds. To this end, we are focusing on several areas:
First, semantic perception through the fusion of large and small models, building maps with semantic understanding. For example, understanding natural language commands like "go to the door" requires semantic comprehension to complete tasks.
Second, real-time mapping and the construction of 3D semantic maps.
Third, hierarchical reinforcement learning for autonomous obstacle avoidance and navigation.
Fourth, task generalization combined with language models, such as instructing a robot to find a chair in a room—a significant challenge.
With hardware and brain capabilities, our ultimate focus is on solutions. After extensive work, we believe solutions must provide various tools, including rapid deployment tools, AI detection tools, human-robot collaboration tools, and traditional SDKs for developers—all essential for robot deployment. In practical applications, we identify four relatively mature scenarios:
First, industrial inspection, due to the urgent need for machines to replace humans.
Second, security patrols.
Third, logistics and last-mile delivery. Hotel robots cannot leave hotel floors, but to operate in residential areas or campuses, hardware forms must evolve beyond traditional wheeled designs, and brain capabilities are necessary beyond fixed delivery routes.
Fourth, education and training.
Let me elaborate on two aspects:
First, DW Deploy. How complex is deploying a robot? Look at the image on the top right—a real deployment in a substation. A large 500kV substation may have around 3,000 or more inspection points. Instructing the robot on what to detect and photograph at each point takes over a month. Our goal is to use AI to reduce deployment costs, including:
1. Enabling robots to autonomously build multi-layer maps with features, geometry, and semantics—semantic maps as mentioned.
2. Deploying robots in virtual environments instead of physically guiding them on-site.
3. Utilizing AI to leverage data from previous deployments, enabling automatic generation of most content for new substations, including inspection points and robot paths.
4. Decoupling all tasks from hardware, so deployment functions remain usable even with hardware upgrades or changes. Our target is to achieve deployment for a substation with thousands of points within two weeks.
With such tools, we deliver solutions to clients. Solutions typically include hardware platforms with various sensors (infrared, gas detection, acoustic detection) and a complete workflow supporting task generation, execution, AI detection, and data empowerment for clients.
Let's briefly review a case to understand current robot applications in real work scenarios.
This system, developed by Lenovo, supports coordinated work between robots and drones. We see tasks being set for the robot here. There are many know-how aspects; for instance, accurately reading hundreds of different meter types in energy and power industries using AI is a challenging capability.
Finally, reports are automatically generated. After robots and drones complete their rounds, daily reports are auto-generated. This is one of the most widely applied cases. This year, State Grid introduced policies involving procurement worth billions for such inspection quadruped robots from this single client.
In short, our goal is pragmatic AI. We believe the current issue with robots is practical implementation. Lenovo Group's efforts are focused on three directions:
First, making robots smarter—primarily enabling them to work effectively, improving anomaly detection capabilities, enhancing semantic and logical spatial understanding, and enabling autonomous task planning.
Second, improving efficiency—simplifying deployment processes and using AI for automatic generation, as deployment currently takes too long.
Third, ensuring safety—a critical aspect as robots increasingly integrate into daily life. This requires high-dynamic robust motion control, multimodal safe interaction to prevent harm to humans, and full-stack trust security to prevent hacking. We have initiated new research specifically on safety, emphasizing its importance as robots become part of our lives.
Finally, a brief mention: our team may not be widely known, as Lenovo's computer and server businesses are more familiar. However, in robotics, we received the Shanghai Science and Technology Progress First Prize this year and previously won the Beijing First Prize. We participate in various domestic and international competitions, including AI, SLAM, and State Grid events, achieving excellent results. Our team has also published papers in top international conferences like ICRA and ICCV, particularly in semantic understanding and virtual deployment, with significant achievements and strong academic exchanges and support internationally.
In conclusion, Lenovo Group is also in robotics, and we look forward to collaborating on every step of robot implementation. Thank you!
Comments