WRC 2026: Embodied AI turns from stunts to ROI; domestic chips enter production AI workloads

The review said startup founders at the event were no longer asking when AGI would arrive; they were discussing how much a robot costs, how stable it is and how many workers' wages it can save in a month. On the exhibition floor, logistics sorting was the most watched scenario. A system from Yuanli Lingji, driven by its DM0.5 generalization model, sorted mixed parcels at about two seconds per piece without rewritten trajectories. The startup's co-founder Fan Haoqiang described generalization as a way to spread hardware costs: the same model and hardware had supported four tasks within a month, from cucumber peeling to conveyor sorting. Zibian's WALL-B model showed a robot predicting the position of a label on an unseen box from subtle surface bumps and turning the box over.

Data collection equipment, once a corner exhibit, moved to center stage. Leiphone said the industry's “data shortage” has pushed companies such as Orbbec and Ant Lingbo to combine depth-sensing hardware with a depth model, while tactile sensing was presented as a key source of high-quality interaction data. He Shan Technology demonstrated finger covers that capture contact, force and slip signals synchronized with first-person images. A market has formed for data infrastructure providers with no legacy hardware business, including Guanglun Intelligent and Kaiwang Data, the report said.

Form factors at WRC reflected a split between spectacle and deployment. Quadruped and wheeled machines dominated real-work demonstrations, while humanoid robots remained the exhibition stars. Yunshenchu showed a robot used for inspections in a Swiss nuclear plant corridor, and Zhishen Technology said it had monthly capacity above 5,000 units and annual capacity of more than one million joint modules. Xuanji Power said its inspection efficiency rose 75 percent in targeted vertical scenarios. On dexterous hands, the review found tactile sensing becoming a basic capability, with one hand embedding 1,900 touch points, while commercial volume remained concentrated in six-degree-of-freedom linkage hands and three-finger grippers. The underlying theme, Leiphone said, was that world models are being used as a “physics compiler” that translates high-level instructions into force, deformation and motion commands.

Separately, QbitAI reported that SenseTime's AI infrastructure business helped HiDream.ai move its video-generation business to domestic chips. The project covered model inference, performance optimization, output quality and toolchain adaptation. SenseTime said LightX2V multi-card parallel optimization achieved a 93 percent speedup for DiT-model video generation on domestic chips, and multi-card output consistency was improved by re-injecting face features. The company also built a unified hardware abstraction layer supporting zero-cost migration across more than 10 heterogeneous chips, and assigned forward-deployed engineers to work inside the customer's workflow. HiDream.ai's products cover more than 100 countries and serve more than 50 million professional users and over 40,000 enterprise customers, according to the article.

QbitAI also reported that Zhipu AI launched and open-sourced GLM-5.3-Flash (320B-A18B) on Aug 26, its first native multimodal model in the GLM-5 series. The model scored 57 on the Artificial Analysis Intelligence Index, the same as Anthropic's Claude Opus 4.8. Before the official launch, it ran anonymously on OpenCode and OpenRouter and logged 62 trillion token calls; SenseTime said all traffic was served by domestic chips. In the same hardware environment, end-to-end service performance tripled from the initial baseline, and hardware efficiency and per-token cost reached parity with mainstream Nvidia GPUs, according to SenseTime. The company's heterogeneous mixed-inference approach delivered 1.25 times the price-performance ratio of an Nvidia H-series system and about 2.5 times the token output of a homogeneous domestic setup at equal cost. Its Token Factory served 2.42 trillion tokens a day in July and plans to exceed 10 trillion daily by the end of 2026.