Beijing Unveils 2026-2028 Action Plan for Embodied Intelligence Robotics Industry

The Beijing Municipal Bureau of Economy and Information Technology announced an action plan for the embodied intelligence robotics industry for 2026-2028 on November 11, 2026. The plan aims to foster at least five internationally competitive humanoid robot manufacturers and at least 10 manufacturers at an advanced domestic level by 2028. It also targets 25% of national finished-robot production capacity in the Beijing-Tianjin-Hebei region and a regional supply rate of more than 60% for key components.

Beijing's targets cover not only the number of finished-robot manufacturers by 2028 but also 25% of national finished-robot production capacity in the Beijing-Tianjin-Hebei region and a regional supply rate of more than 60% for key components. The plan is designed to measure company development alongside regional manufacturing and component supply. Whether the collection of more than 1 million hours of real-world data, dedicated computing resources exceeding 50,000 PFlops and real-world evaluation systems advance together will be a concrete focus in assessing the plan's effectiveness.

A notable feature of the plan is that it places robot makers, models, training data, computing resources, evaluation and safety governance within the same policy framework. Its targets divide finished-robot manufacturers into two tiers — at least five and at least 10 — and set numerical goals for regional finished-robot production capacity and component supply rates. Rather than focusing only on fostering companies, the plan appears designed to assess manufacturing scale and the regional foundation for component supply in parallel. On the technology side, the plan supports VLA, world models and reinforcement learning without limiting support to a single approach, while treating real-world data and a dedicated computing cluster as separate measures. The target of collecting more than 1 million hours of data across more than 100 scenarios indicates the scale of data supply intended to connect model development with real environments. However, the plan text does not specify how data-collection scenarios will be selected or how the data will be evaluated and used. The actual development support provided by a computing resource exceeding 50,000 PFlops cannot be measured without also considering how it is allocated and operated. The proposal for a real-world test site that handles multiple robot platforms points to an approach in which model evaluation is not limited to a specific machine. One issue is how an evaluation method in which users set the tasks will ensure that different platforms and tasks can be compared. The Beijing-Tianjin-Hebei targets of 25% of national finished-robot production capacity and a regional supply rate of more than 60% for key components could measure links between manufacturing and component supply in the region, rather than merely the number of companies. But the scope used to calculate the supply rate and the definition of actual production capacity remain points for future scrutiny. The next points to check are the standards for measuring progress toward each target, the operating arrangements for dedicated computing resources and data infrastructure, and the certification method for real-world evaluation. The plan also calls for safety governance across the entire life cycle, but the information provided does not show which departments will be responsible at each stage or how evaluation results will inform adoption decisions. The plan's effectiveness will depend not only on meeting targets for manufacturers and infrastructure scale, but also on how data, evaluation, supply chains and safety management are connected in real-world operations.