The research team has proposed the first embodied agent framework, RoboFoundry, which formalizes agent systems into self-evolving systems. In this framework, the parameters of the base model remain frozen, while evolution occurs in the context, skill systems, and execution interfaces surrounding the model. Additionally, with the help of shared semantic interfaces, it can be reused across different robots. RoboFoundry employs a dual-loop evolutionary process of execution—reflection—repair—promotion. It first identifies capability gaps responsible for execution failures, then makes targeted modifications and validations, ultimately consolidating these into general-purpose capabilities. It significantly enhances the performance of multiple base models on embodied task benchmarks, narrows the performance gaps between bases, and achieves cross-robot, cross-task zero-shot transfer and post-deployment online evolution in real-world robot scenarios, providing a new pathway for the evolution of embodied AI.