The core challenge of embodied intelligence has shifted from competing over data and parameters in foundational models to the practical deployment of robots and their ability to consistently complete tasks. Silicon Valley generally believes that the current key lies in solving the 'last mile' problem of transitioning from 'can do' to 'reliably completing tasks.' The Cao Ting team at Tsinghua AIR in China recognized this direction earlier and launched the Domain Transformation (Z-Trans) project in 2026, extensively open-sourcing related technologies and proposing the 'Embodied Self-Evolution' paradigm. This paradigm emphasizes that physical agents autonomously refine their strategies through interactive feedback in deployed environments, thereby improving task success rates. Drawing on research experience at Microsoft Research Asia and Tsinghua AIR, Cao Ting led the team in building a comprehensive system, including the Zeva contextual causal learning model, the Zetta online execution and error correction loop, and the Z-Infra training-inference infrastructure. The Zeva model enables robots to extract causal relationships from their own interaction trajectories, enhancing capabilities without parameter updates and achieving top performance in multiple benchmark tasks. Zetta realizes online error correction and skill accumulation through a triple-loop mechanism, with task success rates continuously improving as interaction counts increase. Z-Infra supports full-cycle deployment and data flow, driving the formation of a virtuous cycle where 'interactions generate data—data accumulates experience—experience enhances capabilities—capabilities enable more deployments.' Leveraging four layers of barriers—team and ecosystem, full-stack technology, real interaction data, and scenario compounding—Domain Transformation is propelling embodied intelligence from the pre-training era to the self-evolution era, transforming robots from 'delivery machines' into 'delivery vehicles for sustainably growing productivity.'