Expert Insights on Robotics
On July 20, the 2026 World Artificial Intelligence Conference (WAIC) reached its final day in Shanghai. If asked what was the "hottest" at this year's exhibition, the answer is undoubtedly robots. Hundreds of humanoid robots competed on the same stage, and the embodied intelligence exhibition area was always packed with crowds.
Unlike previous years that focused mainly on "showing off skills", all new products this year emphasize mass production and practical work. As robots can fold clothes and cook, inspect and maintain, and interact and collaborate naturally with humans, and as the coordination between the "brain" and "cerebellum" moves from the laboratory to the production line, this WAIC declares to the world with a full field of steel figures—the "Year of Practical Action" for embodied intelligence is being written.
What can robots do? At this WAIC, various manufacturers have handed in their "answers".
Industry First: Robots Collectively "Enter the Factory"
The most intuitive change at this WAIC is the significant reduction in martial arts performances and dance choreography, which once occupied the center stage, replaced by real operational scenarios such as de-palletizing, material loading, and assembly. Embodied intelligence is completing the transformation from "demonstration-oriented" to "value-oriented". In this transformation, industrial scenarios have taken the lead in submitting their answers.
Leju Robotics
At this WAIC, Leju Robotics, with "Embodied Intelligence, Chinese Solution" as its core theme, officially released the fully localized and deployable solution, building a four-layer fully autonomous and controllable technology system of "data - brain - cerebellum - body".
There were almost no static prototypes at the booth—three mass-production-level robots for carton de-palletizing, plastic box de-palletizing, and small parts loading each performed their duties, replicating the factory production line with 1:1 high precision. Ren Guangjie, Assistant Vice President of Leju, revealed that this solution has been in trial operation in multiple manufacturing factories in the Yangtze River Delta for 3 to 4 months.
The data is more convincing: the carton de-palletizing solution was implemented in the Haichen warehouse, operating stably for 8 consecutive hours with a comprehensive operation success rate of 95.8%; the plastic box de-palletizing composite operation solution was adapted to the production lines of FAW and JAC Group, with a handling success rate of 92%; the small parts loading solution was implemented in the Zhaofeng Electromechanical production line, with a grasping success rate of 94.3%. The overall localization rate of the Kuafu humanoid robot has exceeded 95%.
JAKA Robotics
JAKA Robotics globally debuted the JAKA K1-25 heavy-duty humanoid dual-arm at this WAIC. The product has a single-arm weight of only 25kg, a maximum payload of 30kg, a combined payload of 50kg for dual-arm collaborative operation, and a repeat positioning accuracy of ±0.1mm.
Its core breakthrough lies in the independently developed APEX series high-torque planetary joints, achieving full-link autonomous and controllable drive, force control, and communication. This model can be equipped with a micron-level positioning chassis to form the Kargo wheeled humanoid robot. In the "Modern Era" exhibition area, it jointly built a new energy vehicle (NEV) motor collaborative assembly system with Huayu Electric, requiring only 10 seconds to complete the high-precision locking of 0.5mm fine-thread screws.
Also exhibited, the JAKA π ("π Zai") is only 1.22 meters tall and weighs 42kg net, making it one of the most compact humanoid robots in its class. It is equipped with JAKA's exclusive "brain + cerebellum" integrated architecture, targeting four major B2B scenarios: higher education, new commerce, pan-entertainment, and health care, and supports VR full-body teleoperation interaction.
Both products are underpinned by JAKA's independently developed APEX integrated joint module and the unified Coboπ EI agent platform, achieving core technology reuse and building a "full-gradient, full-scenario" embodied intelligence implementation system covering industrial heavy-duty and commercial services.
Geek+ Technology
The general-purpose humanoid robot Gino 1 made its public debut, adopting a "wheeled chassis + humanoid upper body + three-finger dexterous hand" configuration, mainly targeting warehouse picking scenarios. In the on-site demonstration, Gino 1 can accurately grasp cross-category items such as inflated potato chips, transparently packaged bread, and plush dolls. The whole-body parallel control with hand-foot coordination reduces the single-task operation time by about 30%, and mass production is expected in the third quarter of 2026.
During the same period, Geek+ released the unified embodied intelligence framework Gravity and its core module, the Gravity 4D embodied model. This model introduces 4D physical representation (spatial structure and motion information), improving the task success rate from 73.73% to 78.62% in the LIBERO-Plus benchmark test. The company simultaneously announced the "one core, dual engines" layout, relying on a global logistics network covering more than 40 countries and over 1,700 projects to build a data flywheel, and launched the GINO ECO open ecosystem to promote technology commercialization.
Commercial Breakthrough: From Supermarkets to Homes
Industrial scenarios have verified the "capability" of robots, but the imagination of embodied intelligence goes far beyond factories. At this WAIC, robots are moving from workshops to supermarkets, hotels, restaurants, and even family living rooms.
Yuan Dian Zeroth
"Yuan Dian Zeroth", a family embodied intelligence brand under Lexiang Technology, turns its attention to daily family corners—elderly care, pet detection, and children's companionship. With the theme "Embodied Intelligence Enters Real Family Scenarios", it showcased two main products in the Tencent Cloud Ecosystem Exhibition Area:
The M1 family small humanoid robot is the core model, focusing on high-frequency real family scenarios such as elderly home care, fall detection, mobile safety patrol, children's companionship and interactive learning, and pet behavior monitoring.
The W1 family tracked robot adopts a tracked structure, which can adapt to complex terrains such as grass, gravel, and slopes. It supports capabilities such as autonomous navigation, following, carrying, and remote scheduling, with a maximum load capacity of about 50kg, suitable for scenarios such as family carrying, outdoor accompanying, and camping services.
In terms of underlying technology, Yuan Dian Zeroth adopts a full-stack embodied intelligence technology route, covering a full-stack layout of six technical points: body structure, embodied general model, VLA, joint modules, lightweight robotic arms, and consciousness bionic models. It possesses full-link capabilities of edge-side perception, cloud storage, real-time video access, multimodal retrieval, and AI (Artificial Intelligence) analysis. During the conference, Guo Renjie, founder and CEO of Yuan Dian Zeroth, was invited to participate in the "AI Fusion · Embodied Intelligence Industry Ecosystem Forum" to share insights on commercial implementation. The company also jointly released the "Global Embodied Intelligence Technology Innovation and Application White Paper" with Tencent, Sullivan, and six embodied intelligence enterprises.
The idea of Yuan Dian Zeroth is very clear: family robots cannot rely solely on single-point functions, but require joint support from cloud intelligence, data management, AI (Artificial Intelligence) agents, application ecosystems, and after-sales services. This is precisely the core logic of Yuan Dian Zeroth's continuous deepening of cooperation with Tencent Cloud around the OpenClaw interface, embodied model iteration, and global cloud infrastructure.
RealMan Intelligence
RealMan debuted the RealBot-S2 and RealBot-L2 wheeled humanoid robots at the conference. The two new products share the same technical foundation of full-stack independently developed joint modules, GLN platform service network, and real-scenario data closed-loop, complementing the spatial adaptability of robots for "real work" from the two dimensions of "folding" and "lifting".
The RealBot-S2 can move laterally at 90° with an extremely small turning radius of 413mm, and can squat down to pick up items on the ground after its legs are folded; the RealBot-L2 has a vertical working range of up to 2.1 meters and a lifting load capacity of ≥60kg.
RealMan demonstrated the cross-city teleoperation capability of the GLN remote operation network on-site—the operator in the Beijing Humanoid Robot Data Training Center controlled the robot at the Shanghai booth in real time, accurately completing actions such as warming cups, adding tea, pouring water, and serving soup with millisecond-level latency.
Qianxun Intelligence
Qianxun Intelligence focuses more on the "embodied brain" and publicly demonstrated real-machine long-horizon composite tasks such as living room storage and desktop organization for the first time at this WAIC.
After receiving the "organize the living room" command, the Moz1 robot, equipped with the independently developed Spirit v1.6, autonomously completed the entire process of environment recognition, task decomposition, and classified placement. In the middle, an audience member threw a ball of tissue paper. Moz1 temporarily stopped to re-plan the route, threw the tissue ball into the trash can, and then returned to the original task. In this demonstration, Moz1 not only continuously perceived item positions, task progress, and environmental changes in an unstructured environment, but also dynamically adjusted subsequent actions based on the execution status, demonstrating strong long-horizon reasoning ability and anti-interference. Throughout the process, the robot's real-time thinking and reasoning processes were synchronously displayed on the screen.
The new generation of commercial service robot Moz2 from Qianxun Intelligence also made its debut, targeting scenarios such as supermarkets, hotels and tourism, and office buildings. It is equipped with 32 degrees of freedom, an omnidirectional steering wheel chassis, and a foldable leg structure.
Tianlink Robotics
Tianlink Robotics fully exhibited its full-chain product matrix from core components to multi-scenario terminal machines this time:
Z1 Tianzhen Series New Bionic Humanoid Robot: 1.67 meters tall, weighing 45 kg, with up to 40 degrees of freedom across the body (excluding dexterous hands). It features an innovative structure with a 5-DOF waist and 9-DOF single arm, capable of large-angle rotation and splits.
L3 Tianling Series New Bionic Humanoid Robot: 1.38 meters tall, weighing 37 kg, with super flexibility, 30 degrees of freedom across the body, and 3 degrees of freedom at the waist. It is equipped with integrated joint modules, with a peak torque of up to 145Nm, and features independently developed micro-integrated joints, with a single-arm rated load of 1-2kg.
Core components: It brought key components such as independently developed integrated servo joint modules, harmonic reducers, and super-cycloid joints. Relying on the full-chain product layout, it creates an integrated solution covering the two core scenarios of intelligent manufacturing production line operations and commercial service exhibitions.
Upstream Breakthrough: Behind-the-Scenes Roles Gradually Emerge
The answers handed in by machine manufacturers on the front line are inseparable from the strong support of the upstream industry chain. At this WAIC, a significant change is that the booth density of companies in basic models, data collection, etc., has significantly increased. Dexterous hands, joints, data, and models—these behind-the-scenes roles have also stepped onto the front stage.
Yuequan Bionics
Yuequan Bionics appeared at this WAIC with full-matrix innovative achievements. Relying on the internationally pioneering "bionic tension-compression body robot theory and technology", it brought a series of bionic embodied intelligence products from core components to complete machines, forming a complete industrial closed-loop of "core components - complete machine manufacturing - industry solutions":
The YingShou Y-Hand M2 has 38 ultra-high degrees of freedom, firmly ranking first among similar products globally; the whole hand gripping force reaches 330N, which is more than 6 times that of traditional rigid dexterous hands; the five fingers close in just 0.2 seconds, with a reaction speed close to that of human hands; the repeat positioning accuracy of a single fingertip reaches ±0.04mm.
The XinShou X-Hand M1 focuses on industrial-grade high reliability, featuring human-like hand size and segment proportions. It weighs only 299.7g, has 26 degrees of freedom, and a five-finger gripping force of 200N, capable of achieving a 50kg whole-hand lifting load.
The wheeled humanoid robot Bowen W-Bot 2.0, with the world's smallest chassis, has a chassis area of only 0.238 square meters. It adopts a bionic hip-knee-ankle triple-folding structure, with an operating height extending from the ground to about 2.2 meters. It can flexibly shuttle through compact spaces to complete high-precision operations.
Yingfu Robotics
As a start-up established in July 2024, it appeared with hardcore technology. The company is deeply engaged in the cycloid joint module, a "new generation" technology route—previously mostly used in the heavy industry field, which can withstand impact loads and alternating loads, adapt to harsh environments, and has advantages such as long life and strong impact resistance.
At this WAIC, the industry's first active cooling joint module launched by Yingfu Robotics became the biggest highlight. Heat dissipation is a common pain point in the industry—under continuous intense movement, the joint temperature of humanoid robots will soar to 90 or even over 100 degrees within ten minutes. Yingfu integrated the cooling solution into the joint module, which can quickly drop the motor temperature to about 30 degrees. It solves the temperature rise problem from the source, realizing stable operation of the joint at high power for a long time.
The simultaneously released joint selection platform can achieve rapid selection and custom design by inputting parameters. In the first half of this year, the models of Yingfu joint modules expanded from 2 to more than 10, and supply relationships have been formed with multiple well-known enterprises.
Dexterous Intelligence
Dexterous Intelligence brought its full-stack independently developed dexterous operating system, covering five dimensions: hardware, perception, data, model, and scenario. Core products include a full series of high-DOF dexterous hands, optical-inertial motion capture, and teleoperation real-machine data collection equipment. On-site, it presented four mature implementation scenarios: industrial inspection, biochemical laboratories, new retail, and scientific research education, and completed the adaptation and practical operation of multi-brand collaborative arms and humanoid arms.
The differentiated positioning of Dexterous Intelligence lies in: it is not a single dexterous hand supplier, but a provider of human-like dexterous operation solutions that connects the full link of "hardware - perception - data - model - scenario". The company has established a joint venture with Shanghai Mechanical and Electrical Equipment Co., Ltd. under Shanghai Electric to layout the intelligence of heavy equipment, and reached a strategic cooperation with China Southern Power Grid to promote the integrated application of power inspection and operation. At this stage, it has completed the industrialization implementation of L1 single deterministic tasks and L2 continuous deterministic tasks, and continues to tackle the multi-task generalization ability in complex environments.
The "answer" from the upstream reveals a deep trend: the competition in embodied intelligence is extending from the comparison of complete machines to a comprehensive contest of underlying capabilities such as joints, dexterous hands, data, and models.
Data and Brain: The Decisive Factor in the Next Round of Competition
Technology is growing, scenarios are expanding, and mass production is making collective efforts, but all manufacturers are competing for the same commanding height—data.
Li Sen, partner at RealMan Intelligence, frankly stated: "The future game point of model iteration is high-quality real-machine data." The company has built data training centers in Beijing, Changzhou, and other places, storing over 10,000 hours of high-quality real-machine data. Every real operation accumulates real-machine data, feeding back into model iteration.
Leju Robotics participated in the co-construction of more than ten professional humanoid robot training grounds nationwide, continuously supplying massive real-machine interaction data. Jointly with Ant Lingbo, based on the LingBot-VLA2.0 large model, it created an exclusive solution for logistics sorting. With only 300 pieces of real-machine data and 1 to 2 days of fine-tuning, scenario adaptation can be completed, and the operation success rate has exceeded 80%.
The Spirit v1.6 embodied foundation model independently developed by Qianxun Intelligence adopts an architecture that deeply integrates VLA and world models, and has topped the global ranking in the RoboArena public evaluation.
The StepStar embodied foundation model completed the extreme challenge of 6 robots working collaboratively and uninterruptedly for 15 hours.
Conclusion
At the WAIC 2026 site this year, the performance of robots was indeed inspiring, with precise and neat movements, and smooth and natural interactions. However, behind the hustle and bustle, there is still a regret—almost all robots were "welded" to their own workstations. They are perfect on their own bodies, but if they were given a different shell, could the same model still complete the work?
This involves a threshold that the industry has not yet crossed—generalization.
Currently, most embodied intelligence companies still adopt the customization path of "model binding to the body", where algorithms and hardware are deeply coupled, resulting in the inability to freely transfer capabilities between robots of different forms and brands. Thus, a seemingly simple but essential question is pushed to the forefront: when can we have a universal foundation large model, so that robots, regardless of their height, weight, wheeled or legged form, can share the same "brain", draw inferences from one instance in unfamiliar environments, and seamlessly migrate to a brand-new body?
The answer to this question may be farther than we imagine, yet also closer. Farther, because it tests not only the generalization ability of algorithms but also the multi-dimensional breakthroughs in data ecosystems, simulation platforms, hardware-software decoupling, and standardization systems; closer, because at this WAIC, data training centers, real-machine data feedback, cloud collaboration, and open-source ecosystems have taken initial shape, and the independent evolution of upstream joints, dexterous hands, and models is paving the way for "decoupling". When the industry no longer insists on making one machine learn everything, but enables all machines to learn a universal "learning method", embodied intelligence will truly move from "specialists" to "generalists", and from "performances on workstations" to "partners in life".
The answer regarding generalization has not yet been fully written, but WAIC 2026 has already written the most solid prologue for it with a full field of steel figures and real operations. The future is here, waiting only for the light of general intelligence to illuminate the path forward for every robot.
Data sourced from publicly available information on the internet. Corrections and additions are welcome. This article represents only the author's views and does not constitute any investment advice.