UBTECH Recreates Actual Production Lines 1:1 at WRC

Customer Factory Replicated on Site

At the World Robotics Conference (WRC), which opened on August 22, 2026, UBTECH (Youbi Xuan) showcased a 1:1 replica of a customer’s production line as a way to demonstrate its humanoid robot technology. According to a report by Quantum Bit’s Tian, Yanlin, the company set up approximately 10 industrial humanoid robots—Cruzr S2 and Cruzr Y1—to continuously perform tasks found in actual factories, such as loading and unloading automotive parts, palletizing and depalletizing, and sorting logistics items. These operations ran continuously from the venue’s opening at 9 AM to its closing at 5 PM.

Dr. Jiao Jichao, UBTECH’s Vice President, told Quantum Bit that when customers choose a robot manufacturer, they prioritize not benchmark performance, but hardware stability, scenario understanding, and the operational capability of the solution. He pointed out that completing a one-off task in a lab is merely a stress test of algorithms; true verification comes from continuously completing thousands of operations in a factory environment.

At the exhibition booth, the Cruzr S2 was shown handling the loading and unloading of automotive machined parts and sheet metal components. Two robots were observed cooperating to transport parts over 1 meter in size. Dealing with disturbances like incoming part size deviations, calibration drift, and part deformation, they achieved a final positioning accuracy of less than 1mm. In a small-item sorting task for logistics e-commerce, the robots were shown picking cosmetic boxes of different specifications from mixed location boxes and placing them onto a Demag automatic polediing wall supply station, achieving an average picking cycle of up to approximately 1,100 items per hour.

Simultaneous Deployment of Three Product Lines

At this WRC, UBTECH did not just display a single product, but simultaneously showcased three product lines targeting industrial, commercial, and home consumer markets. For the industrial line, the Cruzr S2 and Cruzr Y1 performed real-world tasks in automotive manufacturing and logistics. The commercial line featured the all-new Walker C1, which demonstrated functions for reception, entertainment, and intelligent tutoring. For the home consumer market, the ultra-humanoid robot U1 was introduced, equipped with the emotion-aware large language model “Resonance-LM,” focusing primarily on emotional support and companionship.

The simultaneous display of three lines was not merely a SKU expansion; it clearly highlighted the divergence in hardware design philosophy between “dancing robots” and “worker robots.” According to the Quantum Bit article, dancing robots are specialized for dynamic performance, stand roughly 1.2 meters tall, are not equipped with vision or force sensors, and use cost-effective planetary reducers in their joints. In contrast, worker robots feature bodies over 1.7 meters tall, are designed for sustained load and long-duration continuous operation, and employ harmonic drives in their joints, which are superior in rigidity and fatigue characteristics.

Clear differences also exist in computational capability. Dancing robots only need to execute pre-set trajectories, so chips like the Rockchip RK3588 are sufficient. Worker robots, however, must perform real-time environmental perception, force feedback, task decomposition planning, and autonomous retry after failure. High-computing-capability chips like the NVIDIA Thor are essential as the execution platform for embedded large language models in this context.

Three-Layer Architecture of Embedded AI

The embedded intelligence UBTECH incorporates into its worker robots is structured in a three-layer system of Understanding, Prediction, and Execution. The first layer is the foundational large language model called “Thinker,” which builds visual, linguistic, and spatial environmental understanding capabilities through the robot’s first-person perspective data. According to public materials, the Thinker model secured 1st place in 9 out of 12 benchmarks for embedded intelligence models under 10 billion parameters, with its largest foundational model reaching 100 billion parameters.

The second layer is the self-developed “Thinker-WM” world model, which is responsible for predicting the outcomes of physical actions. It judges whether an object will fall when grasped, whether a collision will occur during movement, and whether the next action conforms to physical laws. With this predictive capability, the robot can formulate contingency plans in advance for uncertain disturbances in the real world. It can also recognize problems when anomalies like grasp failures occur, adjust its strategy, and retry.

Prior to this exhibition, UBTECH had already moved these solutions from the validation stage to small-scale deployment, gradually expanding scale based on operational effectiveness in customer scenarios. Dr. Jiao Jichao stated that once humanoid robots are deployed at customer sites, it is difficult for a single company to cover the entire value chain. Factors such as the manufacturing of chips, components, and finished machines, scenario deployment, and data feedback will determine the transition from a few trial units to a rollout of hundreds or thousands of units.

Editorial Opinion

This exhibition by UBTECH suggests a shift in the evaluation criteria within the humanoid robot industry. The industry appears to have reached a stage where continuous operational performance on actual production lines, rather than benchmark scores or demonstration videos, directly links to gaining customer trust. The clear separation of hardware design philosophy between dancing robots and worker robots highlights the limitations of the concept of a universal humanoid.

In the long term, the integration of the entire value chain that UBTECH mentioned may determine the industry’s success or failure. Embedded AI models, high-computing-capability chips, precision reducers, and the accumulation and feedback of scenario data—only when all these elements work together does a rollout of several hundred units become feasible. It seems that vertical integration and the speed of ecosystem construction, rather than the superiority of standalone technologies, will become key competitive factors.

Several points remained unverified in this exhibition. To what extent can the sub-1mm positioning accuracy be maintained during large-scale deployment? How broadly can the generalization performance of the Thinker-WM world model be applied across different factory environments? Is the cost-reduction roadmap for Harmonic Drives established? Both technical challenges and cost-effectiveness remain to be seen, and further performance reports are eagerly awaited.

References

• “不是Demo！优必选把客户产线1:1搬进WRC，解锁具身智能真落地路径”, by 田, 晏林 — 量子位, 2026-08-22T14:46:43.000Z (ARR)
• Source URL: https://www.qbitai.com/2026/08/477253.html