Beijing issued a 2026–2028 action plan to substantially upgrade
embodied-intelligence foundation model capabilities. The plan backs multiple
technical routes and their convergence, including visual-language-action models
(VLA), world models (WM) and reinforcement learning (RL). It prioritises
multisource information fusion, physical-space understanding and human
emotion/action recognition to advance models from perception toward cognition —
aiming for “brain” level understanding of physical-law constraints and the
ability to identify non-verbal human actions and expressions. It also targets
real-time inference, cross-ontology transfer and self-learning evolution to lift
multitask generalization, high-reliability execution, fast-cycle operations and
ultra-precision manipulation — the “cerebellum” capabilities. Support for group
coordination, long-term memory, local lightweight deployment and model
interpretability is intended to accelerate real-world deployment. The stated
goal is to bring China’s embodied-intelligence foundation models to
international first-class standard.
