WAIC 2026: China’s Robots Are Starting to Work

At first glance, the World Artificial Intelligence Conference 2026 looked like the world’s largest robot parade. Humanoids greeted visitors, mobile manipulators sorted components, quadrupeds roamed recreated industrial areas, and artificial hands reached for plugs, glass, and delicate components. However, those who only focused on spectacular movements missed the real message from Shanghai: China’s robotics industry is trying to make the leap from impressive individual demonstrations to reliable workflows.

The WAIC took place from July 17 to 20, 2026, in Shanghai. More than 1,100 companies displayed over 3,000 exhibits on more than 100,000 square meters for the first time. According to the organizers, over 300 products were introduced worldwide for the first time. The application show alone brought together more than 300 robots in scenarios from production, trade, everyday life, and entertainment. This scale did not automatically make the fair a proof of market readiness. However, it did reveal where capital, engineering effort, and industrial policy are currently flowing: into so-called Physical AI, that is, AI systems that not only generate texts or images but perceive the physical world and act within it.

The decisive progress at WAIC 2026 was therefore not a single robot. It was a new overall system consisting of specialized body forms, more universal AI models, tactile sensors, training data, charging infrastructure, and pre-series production. The question was no longer just: Can a humanoid walk? It was: Can a robot remain available for eight hours, repeat a task a hundred times, detect malfunctions, cooperate with existing machines, and be maintained economically?

Key Points in Brief

AgiBot introduced four new products for service, research, industry, and manipulation with A3 Ultra, X2 Edu, G2 Max, and OmniHand 3 Ultra-M.
MagicLab presented three new robot forms—MagicBot X1, D1, and MagicDog T1—as well as the Magic-VLA K02 model.
Robbyant, the robotics company of the Ant Group, showcased with LingBot 2.0 the approach of a shared “robot brain” for machines from different manufacturers.
Qianjue Robotics and Fudan University focused on tactile AI: robots should not only see, but also understand force, deformation, and slipping during contact.
Orbbec introduced new hardware that allows humans to collect training data from a first-person perspective, on the wrist, and during fine motor activities.
The transformable Qiyuan T1 questioned the idea that a personal robot must always keep the same body shape.
New training and pre-series platforms in Shanghai show that the real competition increasingly takes place behind the robots: in data, tests, standards, and reproducible manufacturing.

From the Show to Shift Work

Service humanoid and mobile industrial robot working on a modern production line [Image content created with AI] — Not every work robot needs legs: The service humanoid and mobile manipulator take on different tasks. Independent editorial AI visualization.

Robot fairs have lived off choreographies for years. A backflip, a dance, or a short boxing match generates reach and demonstrates motion control. For a paying customer, however, another question is more important: What happens after the hundredth repetition? A production system must be able to handle displaced objects, dirty sensors, fading battery power, changing light, and incomplete information.

It was here that the visual language of the WAIC changed. In a replicated production line for electric vehicles, robots worked on batteries, lights, wiring harnesses, gearboxes, and horns. A system equipped with visual and tactile sensors picked up components, tightened screws, connected plugs, and then checked the lighting. In another demonstration, six robots spent 15 hours together building a model of the Great Wall of China from 80,000 bricks. The task was playful, but the technical core was serious: distributing roles, recognizing work progress, and adapting one’s own actions to the actions of other machines.

The trade fair itself also became a testing ground. According to its own statements, AgiBot deployed more than 60 robots at various event locations. They guided visitors, provided information, and assisted with presentations. Such deployment is not yet proof of a profitable long-term operation. However, it generates real data on navigation, communication, charging plans, and human behavior in a complex environment.

AgiBot A3 Ultra: A Humanoid with a Duty Roster

Among the premieres, the A3 Ultra played a key role. The fully grown humanoid is 1.74 meters tall, weighs 60 kilograms, and has 51 actively driven degrees of freedom. Each arm is said to be able to carry up to five kilograms. AgiBot mentions up to eight hours of combined operating time and supports direct charging, battery swapping, as well as autonomous docking to a charging station.

For perception, the robot combines 3D LiDAR, RGB-D cameras, fisheye cameras, and binocular vision. GPS, RTK, and UWB are intended to enable position determination in different environments. For overarching AI processing, AgiBot uses NVIDIA Thor. This sounds like a long list of technical components, but it is precisely this list that demonstrates the product’s ambition: the A3 Ultra is not meant to be a stage robot, but a system for reception, visitor guidance, commerce, and other public services.

Availability is crucial. “Up to eight hours of combined operating time” does not necessarily mean eight hours of autonomous work under full load. Operating time depends on movement, computing load, payload, and breaks. Nevertheless, AgiBot shifts the discussion in the right direction. A service robot needs not only mobility but also an energy and fleet concept. More about this bottleneck is explained in our background on the battery life of humanoid robots.

G2 Max: Why the better work robot can have wheels

AgiBots G2 Max is possibly more economically interesting than the bipedal star. The robot moves on an omnidirectional base, has force-controlled arms, and can adjust its working height. It is designed for material handling, palletizing, machine feeding, and other physically demanding repetitive tasks.

The design follows a sober logic. Legs make sense when stairs, thresholds, or human-shaped outdoor spaces need to be navigated. On a flat factory floor, however, they cost energy, control effort, and maintenance. A wheel chassis is more stable, can carry larger batteries, and is easier to secure. The humanoid idea is retained in the upper body: two arms, flexible reach, and tools for human workstations. The lower body is adapted to the actual task.

AgiBot demonstrated G2 systems that fetch materials in tablet production, supply testing facilities, and return finished products. This is more important than a single gripping demonstration because it describes a closed process. However, whether this process is economical can only be assessed with cycle time, error rate, intervention rate, and maintenance costs.

MagicLab builds not a robot, but a family

MagicLab also moved away from the idea that a single machine has to solve every task. The company showcased the bipedal MagicBot X1, the mobile industrial humanoid MagicBot D1, and the lightweight quadruped MagicDog T1. The family was complemented by the model Magic-VLA K02.

The X1 is the universal, human-like platform for mobile tasks and full-body coordination. The D1 combines a stable wheeled chassis with two arms for factory work, logistics, and research. The T1 is intended to carry sensors or additional manipulators and is suitable for inspection, plant security, or short transports on more challenging terrain. According to MagicLab, the D1 is already regularly operating in a smart factory by Dreame.

This division is an indication of the maturation of the market. In an early technology phase, a manufacturer often tries to focus all attention on a flagship. In a product phase, variants arise for different cost structures and usage profiles. AgiBot and MagicLab thus demonstrate the same strategic insight: The shared platform can be universal, the body does not have to be.

Magic-VLA K02 and the Search for Longer Workflows

The Magic-VLA K02 model is intended to link task understanding with direct action control. MagicLab demonstrated multi-step activities with it: stacking and sealing boxes, folding flexible clothing, and packing a suitcase. Such tasks are more demanding than classic pick-and-place because the state of the environment changes after each step.

With soft objects, the robot cannot determine where to grasp solely from rigid 3D geometry. A shirt constantly changes its shape while being folded. A box may stand crooked, a flap may not latch correctly, and an item may slip while packing. A VLA model – Vision, Language, Action – attempts to integrate perception, a linguistically or semantically formulated goal, and motion policy.

MagicLab reported a success rate of more than 90 percent for the combined stacking and sealing process. This figure is interesting but should not be equated with an independent industry metric. The number and variance of trials, speed, human interventions, and the exact definition of a successful run remain unclear. The article should not ignore such manufacturer claims, but should clearly label them as manufacturer information.

LingBot 2.0: A Brain for Multiple Robot Bodies

Three different service robots process an order together in a modern pharmacy [Image content created with AI] — One AI model, multiple bodies: Different robots share a complete pharmacy workflow. Independent editorial AI visualization. [Image content created with AI]

One of the strategically most important ideas came from Robbyant, the embodied AI-specialized company of the Ant Group. LingBot 2.0 follows the approach of deploying the same model family on different robots. The goal is “one brain, multiple robots”: capabilities should no longer be fully tied to the mechanics of a single manufacturer.

The fair demonstrated this approach in an intelligent pharmacy. Three different robots received random orders, divided the work, retrieved the medications, and handed over the completed order. According to the organizers, a corresponding solution is already being used in branches of the Shanghai pharmacy chain Guo Da.

If a model can actually be reliably transferred between different robots, it changes the robot market. Today, a learned skill is often closely linked to the camera position, joint geometry, gripper, and control of a specific robot. A cross-manufacturer model would need to recognize differences and translate actions into the respective kinematics. This could thus become a software layer that spreads skills more quickly.

The comparison with a smartphone operating system is obvious, but only partially appropriate. A wrong software command not only causes a robot to crash but can also lead to a physical collision. Universal robot AI therefore additionally requires certified safety limits, clear responsibilities, and local control that blocks dangerous model actions.

The sense of touch becomes the decisive competition

Tactile robotic hand inserts a transparent connector into a vehicle light [Image content created with AI] — Seeing alone is not enough: Tactile sensors and force control enable contact-rich precision assembly. Independent editorial AI visualization. [Image content created with AI]

Vision in robotics is highly advanced. Cameras recognize objects, estimate positions, and create spatial models. However, many errors only occur upon contact: a connector is misaligned, a piece of fabric slips, a glass starts to tip, or a screw reaches its intended torque. Without a sense of touch, the robot notices these situations too late or only indirectly.

Qianjue Robotics therefore presented a new VTLA approach: Vision, Touch, Language, and Action are combined in a multimodal model. The demonstrated systems distinguished surface texture, hardness, and elasticity, detected the onset of slipping, and adjusted movements during manipulation. In a demonstration sequence, soft inserts were removed without damage, examined sensory-wise, and reinserted.

The data acquisition gripper XTac UMI G1 combines visual tactile sensors, IMU, high-resolution encoders, and the state of the gripper. This allows image, force, contact, motion, and slip to be recorded synchronously. These data are valuable because videos alone do not show how hard a person presses or when an object is about to slip.

Qianjue also showed a complete hand with finger sensors and electronic skin. The goal is a network of many contact points instead of individual force sensors. For robotic hands, this represents the same developmental step that high-resolution cameras brought to image processing: turning a coarse signal into a detailed spatial perception field.

OmniHand 3 Ultra-M: A hand as a training and work device

AgiBots OmniHand 3 Ultra-M complements this development from the hardware side. The hand integrates 20 active degrees of freedom into a 630-gram housing. It is intended to grasp up to five kilograms, open or close in about 0.3 seconds, and achieve a repeatability of ±0.2 millimeters. Visually operating sensors in the fingertips and a tactile surface in the palm detect contact and object movement.

Noteworthy is the planned dual purpose. The hand is intended not only to operate on the finished robot but also to be used for teleoperation and the recording of human demonstrations. Consequently, data collection and subsequent execution are mechanically more similar. Movements, contact forces, and grasping states need to be transferred less to a completely different end device.

For precise manipulation, torque control remains crucial. A learning model can generate a grasping plan; nevertheless, fast local control loops must prevent motors from exerting too much force on delicate objects or humans.

Fudan shows why a plug is more difficult than a backflip

Fudan University, together with partners, presented a visual-tactile system for assembling vehicle lighting. The robot picked up screws and connectors, inserted them, monitored forces, and finally checked the function of the light. The sensors used are expected to provide around 40,000 perception points per square centimeter. The university reports more than 10,000 hours of collected interaction data.

Such an assembly task seems unspectacular, but it is technically highly relevant. Components have tolerances, cables move, connectors can get misaligned, and screws can damage a thread. A purely visual system does not reliably detect whether a connector is snapped in with the correct force. The tactile feedback makes it possible to correct the path in contact and assess the achieved state.

The example shows why the industrial future of humanoid robots is not determined by walking speed. The greatest value creation often occurs in the hands: when connecting, inserting, sorting, inspecting, and processing. Precisely there, sensor fusion, tactile perception, force control, and an understanding of the workflow come together.

Qiyuan T1: A Personal Robot Changes Its Shape

The Qiyuan T1 placed a different emphasis. The robot, shown publicly for the first time, is supposed to switch between a wheel-based humanoid indoor mode and a four-legged form for grass, gravel, ramps, and steps. A shared transformer-based architecture is intended to control balance, motion planning, and interaction in both forms.

Indoors, wheels are supposed to be quiet and efficient, as well as enable turning in tight spaces. Outdoors, the four-legged configuration offers more ground clearance and contact points. Intended tasks range from companionship to automatic filming during outings or family activities.

The concept is editorially appealing because it challenges the culturally dominant idea of the human-like all-purpose robot. A personal robot does not always have to look like a human. Perhaps adaptability is more important than anatomical consistency. At the same time, crucial product data are still missing: price, actual battery life, weather resistance, weight, safe transformation, and delivery date. T1 is therefore an interesting bet on the future, but not yet a proven mass market.

Training data without expensive robot hours

Another premiere came from Orbbec. The Robot-Free Data Collection Platform includes EGO for head-mounted first-person recordings, UMI for manipulation data, and WristCam for close-up views of the hand or wrist. This allows humans to perform tasks while simultaneously collecting RGB-D, motion, and interaction data.

This addresses a fundamental cost factor. A robot is tied up during data collection, often moves slower than a human, and can be damaged in the event of errors. Wearable data collection allows many human demonstrations to be recorded in parallel. However, the movements must later be transferred to a robot with different reach, strength, and kinematics. “Robot-free” therefore does not mean that no training on the real robot would be necessary anymore.

Orbbec also showcased, together with Robbyant, a LingBot Depth Filter for the Gemini 330 cameras. It is intended to reduce depth noise, fill in missing areas, and improve edges. Particularly relevant are transparent or reflective objects, where traditional depth sensors often provide incomplete data. This creates a seamless chain from data collection through model training to perception on the deployed robot.

The Factory Behind the Robot Factory

Engineers testing humanoid robots, robotic hands, and different platforms in a training factory [Image content created with AI] — The factory behind the robot: Training, testing, and pre-series manufacturing become strategic infrastructure. Independent editorial AI visualization. [Image content created with AI]

The possibly most consequential news of the WAIC was not a robot product. The national and local innovation center for humanoid robots, together with Huawei, presented a new training field. In addition, a pre-series and testing platform built by the center and Shanghai Electric began operation in Pudong.

The approximately 6,800 square meter facility includes modular assembly lines, testing areas, and an intelligent warehouse. It is intended to accommodate various humanoids, mobile manipulators, and four-legged robots. Capacities of up to 2,000 robotic systems as well as 2,000 joint or hand modules per year are specified. More than 400 tests are to cover components and complete systems. Equipment includes hand test stands, special treadmills, and walk-in temperature chambers.

Such an infrastructure closes the gap between the laboratory and the production plant. Young manufacturers can assemble designs, perform stress tests, optimize manufacturing steps, and produce smaller quantities without immediately building their own factory. Even more important is standardization: comparable tests and shared data formats make progress more measurable and make it easier for suppliers to integrate.

China is thus not only building robots but an industrial learning machine. Data flows from real tasks into models, models are tested on different bodies, components are produced in pre-series, and the results flow back into design and training. This pace can create a greater competitive advantage than a single patented joint.

What was truly new – and what was just newly staged

A serious WAIC review must distinguish between premieres and updates. Shanghai Electric’s bipedal Suyuan was already introduced in 2025. At WAIC 2026, the main interest was in the further developed operating concept with a hot-swap system of two batteries. Matrix-3 was already announced in January 2026; during the fair, the focus was on an update of the WAVE model and new partner programs. Keenon presented a new process for hotel laundry, but largely used familiar robot platforms.

Manned mecha constructions, basketball, and acrobatic demonstrations also provided impressive visuals, but only limited information about economic work. A high joint speed proves neither secure collaboration nor reliable object recognition. It is precisely this distinction that an editorial contribution should consistently maintain.

The Open Questions Behind the Impressive Demonstrations

How Autonomous Were the Systems Really?

In many trade show demonstrations, it is not apparent whether a process was fully autonomous, partially remote-controlled, or strictly limited by prearranged conditions. Teleoperation is not a flaw; it can be a useful safety and learning mechanism. It only becomes problematic when human work is presented as autonomous robot performance.

How often did a task succeed?

A success rate without quantity, object variance, and time frame is not very meaningful. For industrial clients, thousands of repetitions matter, not the best video. Manufacturers should in the future disclose success rate, average cycle time, interventions per shift, and safe failure responses.

What does a productive hour cost?

The purchase price alone is not enough. Decisive are depreciation, energy, maintenance, spare parts, control center personnel, software licenses, and downtime. A cheap robot can be expensive if a technician has to intervene constantly. A more expensive system can become economical if it reliably takes over multiple tasks.

How secure are models and data?

Cross-manufacturer AI platforms and wearable data collection create new attack surfaces. Cameras and microphones can capture confidential production or customer data. Manipulation models must not exceed certified force, speed, and collision limits. Cybersecurity and data protection are therefore part of physical security.

Is there service, replacement parts, and reliable delivery dates?

The robotics industry is full of excellent prototypes. A product only comes into being with documented maintenance, training, spare parts supply, updates, and insurance. Especially with young manufacturers, it is unclear how fleets will be supported after three or five years.

Who won the WAIC 2026

The most convincing product toolkit was shown by AgiBot. A3 Ultra, G2 Max, X2 Edu, and OmniHand target different layers of the market and are complemented by real fleet and factory experiences. The strategically most interesting software idea came from Robbyant: a shared model for different robot bodies could spread capabilities more quickly and make hardware providers more interchangeable.

The deepest technical progress was represented by the tactile approaches of Qianjue and Fudan. As soon as robots understand what is happening during contact, tasks become possible that cannot be robustly solved with cameras and fixed paths alone. MagicLab, in turn, showed that a common technology platform can be sensibly distributed across bipeds, wheeled robots, and quadrupeds. Qiyuan T1 delivered the boldest design for personal robotics.

In the long term, however, the training and pre-series infrastructure in Pudong could be the real winner. Products change quickly; a platform that connects data, tests, manufacturing, and standards accelerates an entire industry.

Conclusion: The best robot was not the most spectacular

The WAIC 2026 did not mark a sudden breakthrough towards the universal robot worker. Reliability, costs, and safety are still too insufficiently documented for that. However, the event showed how priorities are shifting. Legs remain visible, but hands, sensors, data, and charging planning are becoming more important. Manufacturers no longer present just a body, but product families. Models are intended to be interchangeable between robots. Training and manufacturing platforms are intended to shorten the transition to series production.

The decisive criterion for the coming years will therefore not be which humanoid jumps the highest. The system that will win is the one that starts punctually on an ordinary Tuesday morning, works reliably for eight hours, handles errors safely, and in the end leaves a lower cost sheet than the process it replaces.

Sources and Further Information

Bewerte den Beitrag hier!

[Total: 0 Average: 0]

Nico Nuss

Website

Author Nico Nuss has been working on mobile computing and automation software since 2001. Drawing on his experience and strong interest in future technologies, he focuses on robotics and AI.