Chinese Robotics Firms Showcase Embodied AI Advances at 2026 World Robot Conference

During the forum, titled "From Model to Productivity," Xinghai's chief scientist Zhao Xing said the industry is moving from single-point breakthroughs to systematic competition involving models, data, hardware, infrastructure, and applications. He introduced three key advances: an upgraded G0.5 foundation model with native action reasoning, a Fast-WAM world model that reduces single-step inference latency from about 800 milliseconds to 190 milliseconds, and G-Fleet, a distributed reinforcement learning system for robot fleets.

Xinghai also announced plans to release G0.5 MAX and launched a reproducibility program that opens model weights, inference interfaces, benchmarks, and fine-tuning tools to global researchers. The company said the G0.5 model has surpassed state-of-the-art performance on eight international benchmarks including LIBERO and RoboTwin 2.0.

On the hardware side, Xinghai introduced three new robot platforms. The flagship wheeled dual-arm robot Nexo features 20 kg payload per arm, an 8-hour battery, and is aimed at e-commerce, manufacturing, logistics, and service scenarios. The company also unveiled a bipedal robot named Kengo for unstructured environments such as energy and inspection, and Lemo, a low-cost dual-arm desktop platform for developers. COO Li Tianwei said Xinghai has delivered thousands of production orders and targets more than 10,000 deliveries next year.

In panel discussions, industry executives discussed commercialization challenges. Wang Kai, CTO of DeMa Technology, said overseas markets offer shorter payback periods and greater acceptance of current robot capabilities, but return-on-investment remains the key barrier. Tian Ruilin from JD Retail said the company plans to deploy one million robots by 2028 using a common base and vertical models. International participants noted that North American companies treat robots as a necessity but face high ROI expectations, while Southeast Asia's low-cost manual work provides natural data collection opportunities.

In a technical panel on world models, researchers debated whether the concept is hype. Yang Shiyuan of DynaRobotics shared data showing that pretraining on millions of hours of human video can actually reduce robot test accuracy, highlighting the gap between understanding and action. Zhao Xing downplayed data concerns, while Yu Chao of Tsinghua said post-training is key but still relies heavily on human intervention. Li Haoxuan of Peking University called for causal reasoning in robots.

Separately, Guiyu Technology, a startup founded by Zhang Fu, a tenured associate professor at the University of Hong Kong, has raised hundreds of millions of yuan in four rounds of funding within half a year, with investors including Alibaba, Yaoju Capital, and Jinjiu Fund. Zhang, a former student of robotics pioneer Li Zexiang and a former industry consultant to DJI, is building "aerial agents" that do not rely on GPS, remote pilots, or pre-mapped environments. Instead, they use onboard sensing, end-to-end control, and world navigation models to autonomously navigate and perform tasks.

In an interview with Quantum Bit, Zhang said the first phase of products will go into mass production by the end of 2026. The products can fly into tunnels, warehouses, under bridges, and forests without manual control. A later phase aims for autonomous reasoning and manipulation, such as cleaning glass curtain walls. Zhang highlighted his team's Real-to-Sim data collection system, which reconstructs realistic simulation environments from a small amount of real flight data, then generates thousands of variations to train models. He said this reduces reliance on costly real-world data collection and narrows the sim-to-real gap compared with pure simulation. He claimed the system can generate a scene of dozens of kilometers in about ten minutes.

Regarding commercialization, Zhang said the industry is still in its infancy and no product has been mass-produced yet, but scale commercialization should come within two years. He cited applications such as inspecting hill infrastructure and power facilities, where human workers face dangerous conditions. He said autonomous flight can improve efficiency several-fold compared with human-piloted drones, as the aircraft can fly at high speed without constant human intervention. Zhang said his experience at DJI taught him how to bring lab technology to production. He argued that investor interest is not only due to his academic title but also the team's ability to execute. He predicted that as autonomy matures, aerial agents will move from specialized professional scenarios into broader consumer applications.