China’s robotics industry is shifting to a new method of gathering training data for embodied AI robots, resulting in a rapid increase in data volume and processing requirements. This transition is significantly boosting demand for cloud computing resources, according to industry sources at the recent World Artificial Intelligence Conference.
While demand from the commercial sector for embodied AI remains smaller than that for autonomous vehicles, it is expanding quickly, said Huang Yang, Deputy Director of heterogeneous computing products at Tencent’s cloud computing division. Huang noted that data collection and cleaning related to embodied AI now constitute a substantial portion of new business. As a result, Tencent Cloud’s computing workload for embodied AI projects has grown four to five times this year.
To handle this surge efficiently, Tencent Cloud has introduced an innovative computing power scheduling model that leverages unused nighttime capacity for data labeling and processing tasks—helping companies reduce their per-unit computing costs.
Many data service providers are shifting focus from autonomous driving and large language model datasets toward embodied AI, explained Ding Zhezhang, co-founder of a company specializing in AI data infrastructure. These providers mainly handle data creation, annotation, crowdsourcing, cleaning, and formatting.
A new approach to data collection is emerging, called “human data,” which involves capturing information from people wearing specialized devices while performing tasks that robots are designed to learn. Instead of relying mainly on simulations or robot-collected data, companies are increasingly collecting this human-generated data.
Ding explained, “Traditionally, training datasets comprised about 10% to 20% real robot data, with the rest being synthetic simulations. Now, the trend is shifting to roughly 80% human data for pre-training and 20% real robot data for fine-tuning.” The amount of human data available has also grown substantially—from a few hundred hours previously to datasets exceeding 500,000 hours, with some experts believing it could reach 10 million hours in the future.
This method of data collection is already being implemented in real-world settings. Ding’s company has set up data collection in factories, where frontline workers wear devices to gather task-related data. Though interest was minimal last year, recent demand has surged so rapidly that they’re struggling to keep up.
The growth in embodied AI data has also driven up data processing needs. Processing human data is much more labor-intensive than handling machine-sourced data, Ding noted, and his company’s platform relies heavily on cloud infrastructure to manage these workloads.
Despite the surge in training data, industry experts believe it will still take time for robots to see a major leap in capabilities. Real-world robot data remains limited, especially in niche applications, making progress slow. Huang compared embodied AI development to autonomous driving, noting both face a “long-tail effect”—initial improvements are rapid, but reaching the final 20% of performance requires significantly more effort.
“There’s still a gap between what people expect from embodied AI robots and what current hardware and AI models can deliver,” Ding said. “The data flywheel, a positive feedback loop where data and models improve each other, has yet to fully start but is expected to become operational within the next two to three years.”