Chinese AI Robotics Startups Race to Teach Humanoids Real Skills

Chinese AI Robotics Startups Race to Teach Humanoids Real Skills

Chinese robotics startups are converging on a critical challenge: teaching humanoid robots to perform human tasks with the fluidity and adaptability that real-world deployment demands. At the World Robot Conference in Beijing last week, several companies revealed their approaches to bridging the gap between laboratory demonstrations and commercial viability.

QJ Robots, a three-year-old startup founded by Tsinghua University automation PhDs, has secured backing from Temasek and claims to have generated more than 100 million yuan ($15 million) in revenue. The company has deployed its AI model to over 100,000 robots, primarily for mainland Chinese enterprises, according to founder and CEO Haichuan Gao.

The startup's focus centers on world models—AI systems designed to predict and simulate physical environments rather than generate text like ChatGPT. Chief Technology Officer Tianren Zhang emphasized that while the company collects data monthly from its large robot deployment base, the variety of data remains insufficient for optimal model training.

Industrial Integration as Competitive Edge

Beijing's strategy encourages industrial giants to develop integrated AI capabilities, creating a manufacturing ecosystem that startups are leveraging. Lumos, a two-year-old company founded by another Tsinghua alumnus, Chao Yu, has partnered with Mitsubishi Electric and established its own data collection center.

This year, Lumos launched NexCore, a platform designed to translate data and AI instructions into executable tasks across multiple robot brands—not just its own hardware. The move reflects a broader industry recognition that interoperability may prove as important as individual robot capabilities.

Ace Robotics, launched just one year ago with SenseTime Co-Founder Wang Xiaogang as chairman, takes a different approach. The company produces wearable sensors to capture precise human movement data, arguing that training on online videos introduces unrealistic elements into AI models. Wang predicts the industry will reach a breakthrough moment by the end of next year, comparable to ChatGPT's transformative launch in late 2022.

Reality Check at the Exhibition Floor

Public reaction at the conference exhibition revealed the gap between current capabilities and commercial expectations. Robots performing tasks like folding clothes or dispensing water moved slowly and stiffly. One visitor waiting for a robot to serve water noted that people remained patient only because a machine was performing the task—a human moving that slowly would face criticism.

Unitree's shares declined after founder Wang Xingxing tempered expectations for near-term humanoid commercialization, underscoring investor impatience with the technology's development timeline.

Why it matters

The race to build commercially viable humanoid robots represents a test of whether China's manufacturing ecosystem and data collection advantages can translate into AI leadership. With multiple well-funded startups claiming different paths to the same goal—teaching robots human skills—the industry faces a critical question: whether markets will wait for the breakthrough moment these companies predict, or demand faster returns on substantial investments.

These details were first reported by CNBC's The China Connection newsletter, written by Evelyn Cheng from Beijing.