(Yicai) July 18 -- China’s robotics industry has embraced a new approach to acquiring training data for embodied AI robots, leading to a sharp increase in both data volumes and data processing needs. This change is fueling rapidly growing demand for cloud computing resources, Yicai learned from industry participants at the World Artificial Intelligence Conference which kicked off yesterday.

Although commercial demand for embodied AI remains far smaller than that for autonomous driving, it is growing rapidly, Huang Yang, deputy director of heterogeneous computing products at internet behemoth Tencent Holding’s cloud computing arm Tencent Cloud, told Yicai.

Demand related to embodied AI data collection and data cleaning has recently accounted for a significant portion of new business, Huang said. As a result, Shenzhen-based Tencent Cloud's embodied AI-related computing workload has surged four- to five-fold this year.

To meet this demand more efficiently, Tencent Cloud has introduced a new computing power scheduling model that utilizes idle computing capacity during off-peak nighttime hours for data labeling and processing, helping embodied AI firms reduce their unit computing power costs.

Many data service providers are shifting from autonomous driving and large language model datasets to embodied AI, said Ding Zhezhang, co-founder of embodied AI data infrastructure company IO-AI Tech. These firms specialize in data generation, annotation, crowdsourced collection, cleaning, processing and format conversion.

Human Data

Behind this trend, a new method of data acquisition is emerging where instead of relying primarily on simulation data or data collected directly by robots, companies are increasingly collecting what the industry refers to as "human data," which is data captured by people wearing specialized data collection equipment while performing tasks that robots are expected to learn.

"In the past, a typical training dataset might consist of between 10 percent and 20 percent real robot data and 80 percent to 90 percent synthetic simulation data," Ding said. "Now, a new approach is emerging which uses approximately 80 percent human data for pre-training and 20 percent real robot data for fine-tuning."

The scale of available human data has also expanded dramatically. Whereas training a model previously relied on only a few hundred hours of human data, datasets can now reach 500,000 hours, with some industry participants believing they could eventually grow to 10 million hours.

This data collection approach is already being deployed in real-world settings. Ding said his company has entered factory environments, where frontline workers wear data collection devices. While there was little market interest in this type of service last year, demand has recently surged to the point where the Shenzhen-based firm is struggling to keep up.

The rapid increase in embodied AI data is also creating significantly greater data processing requirements, further driving demand for computing infrastructure.

Processing human data is much more labor-intensive than processing data collected directly from machines, Ding said. His company's data platform relies on Tencent Cloud's storage and computing infrastructure to support these workloads.

Data Bottleneck

Despite the sharp increase in available training data, industry participants believe it will still take time before robots achieve a major leap in capability.

One reason is that real-world robot data remains indispensable, yet it is still in short supply and data for specialized application scenarios remains scarce.

Huang compared embodied AI with autonomous driving, in that they both face a "long-tail effect.” Progress can be rapid through the first 60 percent, 70 percent or even 80 percent of capability development, but achieving the final 20 percent requires disproportionately more time and effort, with the last few percentage points being the most difficult.

"There is still a gap between people’s expectations for embodied AI robots and what current robot hardware and AI models can actually deliver," Ding said. “The data flywheel, which refers to a self-reinforcing positive feedback loop in which data and models continually improve one another, has yet to take off, but is expected to begin operating within the next two to three years.”

Editor: Kim Taylor