China Launches Multimodal Dataset to Break Embodied Intelligence Bottleneck

People’s Daily Online reports that a key breakthrough has been made in addressing the bottleneck that restricts embodied intelligence from "being able to act" to "understanding human intentions" at the 2026 People’s Network Data Intelligence Partner Conference. The Mainstream Value Corpus Ecological Alliance was officially launched during the event, alongside the release of China’s first "embodied interactive multimodal corpus dataset" by the "AI Interactive Corpus Innovation Laboratory" co-established by People’s Network and Youzu Network.

The dataset integrates People’s Network’s mainstream value corpus with Youzu Network’s digital interactive behavior data, providing core data support with both technological depth and ethical thickness for the development of the embodied intelligence industry. People’s Daily notes that the AI Interactive Corpus Innovation Laboratory, founded in late 2025, focuses on building a safe and reliable multimodal human-computer interaction corpus system to support the advancement of embodied intelligence.

Through collaborative research with the State Key Laboratory of Autonomous Intelligent Unmanned Systems at Tongji University, the laboratory has built the first batch of high-quality interactive corpus datasets with a scale of 100,000 entries based on the Vision-Language-Action (VLA) underlying framework. The VLA framework integrates visual, linguistic and action capabilities, enabling end-to-end mapping from input to machine action execution.

The dataset includes robot action data, facial expression data and emotional intelligence training data, covering three categories and 20 sub-scenarios such as guide services, home services, games and virtual digital humans. It achieves timestamp alignment of actions, expressions and voices through a hardware synchronization trigger mechanism, establishing four annotation systems: action, expression, language and alignment.

Leveraging People’s Network’s authoritative corpus ecology, the dataset places special emphasis on social responsibility, cultural connotation and ethical compliance, ensuring that robots can understand and follow human social logic while completing instructions. This addresses the data bottleneck in embodied intelligence development, which, along with ontology and model bottlenecks, has long hindered the industry’s progress.

As one of the achievements of the Mainstream Value Corpus Ecological Alliance, the dataset will open interface access to research institutes and model manufacturers in the future. During the event, a cooperation agreement was signed between the dataset and Lingyi Wanwu, a representative domestic large model company, to explore collaboration in data usage, model training and overseas services.

Guided by People’s Daily, People’s Network initiated the Mainstream Value Corpus Ecological Alliance to build an open and collaborative platform for mutual benefit, serving as a link connecting government, industry, academia and research sectors. In the future, the alliance will promote the efficient supply, processing and application of mainstream value corpus, while the new dataset will accelerate the application of embodied intelligence in home services, exhibition guidance and other fields, advancing the development of a safe, reliable and human-centered embodied intelligence industry.